When I get off a scoping call, I literally type one sentence. And about 10 minutes later there is a meeting record, an internal overview and a client-ready proposal all sitting in my Notion.

I do not write them. I actually never read the transcript. I do not open the recording to review it.

Look, it might sound strange but I used to spend about 3 hours on a single client proposal. Digging through call notes, copying numbers over, formatting, second-guessing every line of it. Now it is under 10 minutes, and most of those 10 minutes are just me waiting for the meeting recording being transcribed.

What this article offers is a copy-paste ready system. Every file is in here, you don’t have to invent anything, and if you change the names to yours it runs today.

Subscribe to No BS AI

One real copy-pasteable build like this every week.

Subscribe now

What happens if you just paste the transcript in

Most people try the obvious version first and I did too of course.

You download the transcript, you drop the hour-long transcript into a chat and ask for a proposal. Something mediocre comes back. It looks fine. Then you spend 40 minutes rewriting it of course, and you are not really sure why.

You’re questioning your life decisions and asking yourself: “Why am I doing this?”

And usually 3 things are going wrong at the same time here:

• It invents prices. Nobody mentioning a single number on the call, the model just picks one that sounds about right, and now there is a rate sitting in your document that you never agreed to.

• It writes to the wrong place. Say you asked it to save into Notion, and your property is called “Client” while the agent went and guessed “Client Name”.

• It promises work nobody asked for. You mentioned one integration, one system, the model went and scoped four of them, and now all four are sitting in a document and, again, you find yourself with something you never agreed to.

So technically you have a proposal, and practically you have a first draft that you now have to audit line by line, and it’s not gonna be 10 minutes. That is the 3 hours all over again, just with a bit more extra steps in the middle.

And this is normal by the way. Proposals are slow almost everywhere. Loopio runs an annual survey of response teams, and in the 2026 edition, across 1,500+ teams, the average RFP takes 33 hours of work, down from 35 the year before. Loopio sells proposal software, so read their number with that in mind, but the direction matches what every consultant I know says out loud.

This is normal because it’s what everyone does, but we’re not here for normal, are we?

What you get below this line

The skill system that runs the whole sync and creates you 3 pages for each call.

Complete, nothing truncated, exactly the same system runs on my machine.

You get the parts that go wrong as well, because they will. Your first run is very likely to report back a beautiful finished proposal and write absolutely nothing anywhere, and I will show you how to catch that in about 5 seconds instead of the afternoon it cost me.

The only thing I still do by hand after all of this is the pricing.

No code and no terminal anywhere in it by the way. If you can make a folder and paste text into a file, you are good to go.

Quick heads up: a free overview of this build is coming out in 2 days, so stay tuned. If you want my complete personal build, that I use every day, and not just a high level abstract overview, please, follow along.

Step 1: What you need before you start

Three things, and you probably have two of them already.

A call recorder that keeps full transcripts. For instance, I use Fathom. It just sits in the call, records everything, and keeps both a summary and the full transcript. The summary on its own is not going to be enough, mostly because it doesn’t capture the whole set of specific details that you might want to extract from the call.

Notion with 3 databases. Meetings, Projects and Clients. And if you already have some kind of CRM-ish setup in there, you probably have all 3 already just under different names, which is completely fine because you are going to write your real ones down in a second anyway.

An agent that supports skills and connectors. I run Claude. A skill is just a folder with a file in it that says how you do something. And a connector is a simple way to connect your favorite apps (e.g. Gmail, Slack…) to Claude.

Now connect exactly 2 things. Your recorder to read from (e.g. Fathom, Granola), and your Notion to write to. Nothing else really. Every connection you switch on takes up room in the agent’s head, so 2 tools you actually use every day will beat 15 that you turned on once and forgot about.

Apify is just the example in those screens by the way. It looks exactly the same when you go and connect Fathom or Notion.

And then test each one on its own, in a clean chat, before the skill ever goes near them:

• “List my last 3 calls from [recorder, e.g. Fathom].” You should get back real titles and real dates.

• “Create a page called Proposal test in my Notion.” Then go open it and look at it.

And if either one of them fails, fix it now. Debugging a broken connection through a 5 step skill later on is genuinely horrible.

When you connected everything you should see something like this:

Step 1 done.

Step 2: the one file that does all of it

So here is the whole build, and it is one single file.

One skill, 5 steps, and 3 possible outputs depending on what the call was. Every call gets a meeting record no matter what. A call that starts a new engagement gets a project on top of that, and a call that was really about scope also gets an Overview and a Proposal written underneath it.

The 3 page shapes are in there too, near the bottom, and they turn “write a proposal” into a document that comes out looking the same every single time. Which matters more than it sounds. Because when the shape never changes, you can skim a proposal in 2 minutes instead of properly reading it, since you already know what is sitting in position 4 before you even get there.

Later on just swap my names with yours and run it.

Here’s the prompt:

Read more

So you just subscribed or simply passing by, welcome to the No BS AI newsletter!

There are many pieces published on the newsletter, but you don’t know where to start.

Well, let me help you on this.

You probably did not come here to read all of them of course, but you certainly came here because one specific thing is annoying you.

Subscribe now

So here’s the deal: every line below says what you actually get out of the piece, not what the piece is about. Pick the one which title matches your problem and implement it. Almost all of them are actionable, practical guides that will make your life easier.

A few of them are paid, and I’ve marked those of course so you don’t click into a paywall by surprise.

This is how you get leads

How I got 25 leads in 15 minutes, for $0.12 with Claude

How to Find People Who Actually Buy From You - The Overview

How to Find People Who Actually Buy From You - Complete Guide (paid)

This is how you get AI to write in your voice

AI is guiding your voice

How to Treat Claude Like A Game Character - The Overview

How to Treat Claude Like A Game Character - Complete Guide (paid)

This is how you stop repeating yourself every session

SKILLs changed everything

Claude Connectors = Second Brain

This is how you stop wasting the tool

Claude is not Google

My 5 Levels of Claude

The 3 tools I use to grow my AI business

This is how you grow the newsletter

How I Doubled My Substack Notes Engagement

The real reason your Substack isn’t growing (paid)

This is for the week you feel behind

Just my personal reflections on some AI automation and entrepreneurship topics.

Everyone’s Winning at AI

Harsh Truth about Listening to AI GURUS

There’s no competition

You DON’T need a lambo

Own your path and This Path Isn’t for You

When Everything Works and Nothing Works

The part of AI no one talks about

Where I’d start

If you have never built anything with an agent before, read SKILLs changed everything first, then Claude Connectors = Second Brain. Those 2 are the floor that everything else stands on, and both of them are free.

If you already build things and you just want more clients, go straight to How I got 25 leads in 15 minutes.

Subscribe now

Talk soon,

Ilya

I was so tired of giving my AI agent corrections all the time. It’s either stylistic, or length, or content correction. It always says yes, and then the next session it does the exact same thing all over again.

So I type it again. It works for a message or two, and then I open a new chat the next morning or the next hour or so, and I’m back to rewriting half of everything I asked.

It was so frustrating so I realized that there must be a way of creating the perfect settings adjustments for your agent.

So I thought about the game character analogy.

To make the long story short, I found that there are 10 small tweaks that can help you fix this, and none of them need you to install a single thing. So, please, follow along.

Subscribe to No BS AI

One real copy-pasteable build like this every week.

Subscribe now

The problem

The main problem with the current AI agents is that they agree with things they should be arguing with. They write 4 paragraphs when you only wanted one. And they ask permission for something tiny, then that task goes completely off and the agent does something huge without asking at all.

The agreeing part has real research behind it by the way. A study called SycEval gave models a correct answer and then pushed back on it anyway. Across ChatGPT-4o, Claude Sonnet and Gemini 1.5 Pro, 58.19% of the samples moved.

Now most of that movement went toward the right answer, 43.52% of it, while only 14.66% moved toward a wrong one. So they are not simply caving, they are mostly getting corrected.

But once a model flipped, it then stayed flipped 78.5% of the time, which means asking it again in the same chat is not really giving you an alternative opinion. You are asking something that already agreed with you once.

Why writing more instructions does not fix it

There are 2 reasons for why writing more doesn’t get this problem fixed, and both of them boring unfortunately.

The first one is that an instruction in the chat lasts about as long as the chat does. You tell it to be blunt, it is blunt for one message, and then it goes back to normal while the thing you are actually fighting shows up in every message after that.

So you move the instruction into a file. Which is better of course, and most people sit right there, but I’m gonna tell you that files grow. You add a rule after a bad session, you never delete it, then you add another one, and 6 months later you have 400 lines that nobody reads anymore, including the model.

That last part is measured by the way. Anthropic’s own engineering team wrote that models have something called an “attention budget”, and that as the number of tokens in the context window goes up, the model’s ability to recall anything out of it goes down. Their advice is to find “the smallest possible set of high-signal tokens”.

10 numbers is a small set of high-signal tokens. A 400-line rules file is not.

What a character sheet actually is

I decided to still use a file to treat this problem, but without letting it grow indefinitely, and I got the idea of treating the agent as a game character with its own stats. It’s somewhat of a dimmer switch that you gradually adjust to the value you need.

And actually, most AI instructions are light switches. “Always push back on bad plans” is on. “Be concise” is on. There is no in between anywhere, so the model picks its own in between, and it picks a different one every session.

A character sheet turns each of those into a dimmer instead. Each word represented on a scale of intensity from 1 to 10, and one sentence telling the model what a higher number means.

Here is what mine looks like once all 10 of them are set. The shape is the whole point: you can see at a glance which way each dial is turned, without reading a single line of instructions.

As you can see I bumped my Critique up in order to let Claude agree with me less.

One number does the job of 5 huge paragraphs.

What you need

For building something like this you’d need 3 ingredients, and you probably have 2 of them already:

• Somewhere to store the numbers. A file, a note, the project instructions box. Anything your assistant actually reads. This can be a json file for example, see below.

• An AI assistant that reads standing instructions. Claude with a project, the custom instructions box, whatever you already use.

• About 1 hour, and most of that is just you deciding what your numbers should be.

No terminal, no code, nothing to install.

Step 1: write the numbers down

Here is the shape of it. This is just a template you can use in your own project, you get 3 of the 10 stats, so you can see how one row works before you go and write your own.

{
  "attributes": {
    "critique": {
      "value": 8,
      "effect": "Challenge a weak plan before building it. Name the problem, propose the better move, then wait."
    },
    "charisma": {
      "value": "<your number, 1 to 10>",
      "effect": "<what a HIGH number means. Example: answer at length, with examples>"
    },
    "wisdom": {
      "value": "<your number, 1 to 10>",
      "effect": "<what a HIGH number means. Example: look inside a file before overwriting it>"
    }
  }
}

Critique is filled in as a demonstration so the shape is obvious, and the other 2 are yours to write. The complete 10-stat sheet, with the numbers I actually run, is in the paid edition.

Now effect is the important one. That is the only sentence the model really reads, so you want to write it as an order and not as a description. “Look inside a file before overwriting it” tells it what to do. “This stat represents carefulness” or, the one that makes me laugh every time, “be more human” tell it nothing at all.

Step 1 done. You have the numbers locked in, but they mean nothing and nothing reads them yet.

Step 2: tell it what the numbers mean

A number on its own without a specific meaning is just a decoration. critique: 8 means nothing to a model that was never told what critique is.

So the Stats go next to the numbers. Paste this in and edit the list to match whatever settings you picked:

## Stats

You have a character sheet of stats, 1 to 10 each. They are live behaviour
levers, not decoration. Read them and apply every one of them. Higher =
more of the effect.

- Critique. Your pushback bar. Below it, say what is wrong before you
  build anything. At or above it, just build. If I say go anyway, do it
  my way and drop it.
- Charisma. Output length. Low = the result only. High = the full
  explanation with examples.
- Wisdom. Data-loss guarding. High = look at what you are about to touch,
  and copy it first.

And the last line of the Critique block does more work than the rest of it. If you leave out this phrase “if I say go anyway, do it my way and drop it”, you then get an assistant that argues, and then argues about the argument, and you will eventually stop reading its objections.

Step 2 done.

Step 3: put it somewhere it gets read

Now, open a project in Claude, or your custom instructions box, and paste both blocks straight in. The numbers and the manual, together, in the same place.

Then it gets read at the top of every chat inside that project, without you doing anything at all.

One catch worth knowing. Doing it this way means that when you change a number, you edit it in the box yourself. The automatic version reloads it for you at the start of every session, and that is the wiring in the paid edition.

Step 3 done.

In practice

Now, if you want this in your system copy-paste the prompt below into Claude → Set Auto Mode → Launch Prompt → Done.

## Context
Your creating a character sheet for adjusting how you write your responses through sessions.

## Instructions
Your job is to:

- Read and analyze the below article
- Understand what needs to be implemented
- Implement everything that has been said in the appropriate Claude Code folder

## Pasted article
[Paste this article here]

Then, start with one number. Whatever annoys you most this week, for example the “language difficulty level” or the “output prompt length”, find the word for it, give it a number, and paste it in.

Then leave it alone for a couple of days. Do not tune anything until something irritates you twice.

And when something does, move that one number by 1 or 2 for the start and see how it changes.

The paid edition this week has my own sheet in it: all 10 numbers with the values I actually run, the script that loads them automatically at the start of every session, and the 4 ways this whole thing breaks quietly on you.

Subscribe to No BS AI

One real copy-pasteable build like this every week.

Subscribe now

Talk soon,

Ilya

Look, I had this pretty big problem of Claude giving me bs output all the time. It was returning me EM dashes, weird formatting, proposing stylistic structures I have never used. It was so frustrating to tell to edit it in every session.

Then I decided to dive deeper and I realized that just blindly prompting and adding rules upon rules in a file that Claude never reads is a dead end.

So I solved it in a way I haven’t seen anyone doing it before.

What the situation look like right now is this: my AI assistant has 10 numbers attached to it, and when it starts misbehaving those numbers are the only thing I actually edit to put it back on track.

For example I have:

• Critique set to 9

• Boldness set to 6

• Wisdom to 8

• Speed to 5

I’ll explain what those numbers mean in just a second, bear with me.

So those are just raw values, sitting in a file under a claude folder right now. As you probably noticed, I set this thing up like a video game character. I genuinely believe it’s the simplest analogy there is for making Claude behave the way I actually want.

So Claude reads those stats at the start of every single session.

So when it hands me a solution which is too good to be true, or just starts agreeing with everything I say, I do not write another paragraph of instructions and sit there hoping.

I open that file, or better my character dashboard, move the critique stat up by 1, and start a new session.

And that is really the whole thing. 10 numbers instead of a growing wall of rules that nobody reads.

Look at how it looks on my end:

The problem

You give it a correction, it says yes, and then the next session it just does the exact same thing all over again.

And if you ask anyone who uses AI every day you get pretty much the same short list. It agrees with things it should be arguing with, then it writes you 4 paragraphs when you only wanted one, and it asks permission for something tiny, then goes and does something huge without asking at all, which has happened to me more than once honestly.

By the way, the agreeing part has real research behind it. A study called SycEval gave models a correct answer, then pushed back on it anyway. Across ChatGPT-4o, Claude Sonnet and Gemini 1.5 Pro, 58.19% of the samples moved. Claude Sonnet on its own, 57.44%.

Now most of that shift went toward the right answer actually, 43.52% of it, while only 14.66% moved toward a wrong one. So the models are not simply caving, they are mostly getting corrected.

But once a model flipped, it then stayed flipped 78.5% of the time.

When you’re asking something that already agreed with you once, it will probably just agree again.

I noticed that the judgement has to come from outside the conversation.

Why writing more instructions does not fix it

And there are 2 reasons, both of them boring unfortunately.

So the first one is that an instruction in the chat lasts about as long as the chat does. You tell it to be blunt, it is blunt for one message, and then it goes back to normal, while the thing you are actually fighting shows up in every single message after that.

Then you move the instruction into a file. Which is better of course, and most people stop right there, but files grow. You add a rule after a bad session, you never delete it, then you add another one, and 6 months later you have a huge 400 lines file that you’ll never read. I’m talking from personal experience here.

That last part is measured because Anthropic’s own engineering team wrote that models have an “attention budget”, and that “as the number of tokens in the context window increases, the model’s ability to accurately recall information from that context decreases”. Their advice is to find “the smallest possible set of high-signal tokens”.

10 numbers is a small set of high-signal tokens. A 400-line rules file is just noise.

What a character sheet actually is

A character sheet, I’d say, is just a bunch of statistics your AI agent has, basically a dimmer switch that you adjust.

Most AI instructions are similar to light switches though. “Always push back on bad plans” is on. “Be concise” is on. There is no in between, no gradation, so the model picks its own in between, and it picks a different one every session.

A character sheet turns each of those into a dimmer. Just one simple word, just one simple number from 1 to 10, one sentence saying what a higher number means.

Critique 9, for example, means push back very hard before providing any substantial changes.

Boldness 6 means act on your own, but stop at anything you cannot undo.

Charisma 5 means answer at normal length, not an essay.

So 10 of these, all sitting in one small file. Your assistant reads them at the start of every session, and when something annoys you, you go and move one number.

One number does the job of 5 big paragraphs.

What you get from this article

So that was the doctor’s diagnosis. Your AI agents has no settings at all, so it just invents its own every session.

Below the line you get the full build:

• The actual file with my 10 numbers in it, complete, copy and paste.

• The block you put in your instructions so the numbers mean something instead of sitting there just as a fancy decoration.

• The small script that loads them at the start of every session, and the one line of settings that runs it.

• Then my tuning log. Which number I actually moved since I built this, what made me move it, and what I found when I read the file’s history. It was not what I expected at all, and it will probably save you a bunch of tuning time.

This is what my own assistant reads before every session. One small file, a 20 line script, one line of settings. I have not touched 9 of the 10 numbers since I wrote it.

Ok. Enough talk. Let’s build it.

I know a script and a settings file might sound technical. Honestly they are not. One file you copy, one line you paste. And if you never open a terminal at all, there is a version of this for you in step 3.

Quick heads up: I will be publishing a free overview of this build in 2 days, so stay tuned. If you want the complete build, it’s only accessible under the line, then follow along.

Read more

A few months ago I bought a list of 500 people. Real names, real companies, real links. On the screen it looked very good.

Then I sat down to find out which of them were worth a message.

You open the first row. You read the profile. You open the company website. You think for a minute. You decide. If you do it properly, that is 5 to 10 minutes for one person.

At row 30 I was tired. And here is the honest part. After row 30 I was not really checking anything. I was guessing.

A bigger list does not solve this. It only gives you more rows to guess.

This is how most people check a list:

• You open the profile

• You read the short bio

• You look at the last few posts

• You decide by feeling, and in the middle of the list your feeling is tired

But the checking is not the difficult part. You already know what a good client looks like. The problem is that the rules are only in your head. So you are the only person who can use them, and you can only use them for about 30 rows.

Write the rules down one time, and a machine can do the first pass for you. It sorts the list. You still decide.

What you get from this article

This article is the overview. At the end of it you will know:

1. What you are building, in simple words

2. The one field that decides if it works at all

3. The 4 rules that decide if you can trust the answer

4. How the whole thing works, in one picture

5. An empty template of the skill, for you to fill in with your own rules

6. One thing you can do by hand today, in 10 minutes, with no tools

Number 6 is not a small extra. It is the most difficult part of the whole system, and you can do it before you build anything.

Let’s start.

Subscribe now

1. What you are building

You are writing a short file. It says 3 things: who you want, who you never want, and how to put a row into 1 of 3 piles.

The 3 piles are simple. Message this person. Look at this one yourself. Delete this one.

That is all. No dashboard, no tool, no API key.

In Claude this file is called a skill. It has 3 fields: a name, a description, and the instructions.

I know, this sounds boring. Please stay with me for one minute, because the difference between those last 2 fields is the most important thing here.

Step 1 done.

2. The one field that decides if it works

The description is one paragraph. Claude keeps it in every conversation you have.

The instructions are the long part with your real rules. Claude does not read them at all until the description matches what you just typed.

Please read that again, because nobody tells you this part.

That one paragraph decides if the long part is ever read.

So this is what happens to most people. They write 2,000 careful words in the instructions. Then they write “scores leads” in the description and they stop there. They give Claude a list, Claude answers with normal advice about lead qualification, and they think the skill is broken.

The skill was fine. Claude never opened it.

So write the description with the words you really use on a Monday morning. Put the real sentences in it: “score these leads”, “which of these should I message”, “who do I contact first”, “is this person worth my time”. Five or six of them.

Then test it in an honest way. Open a completely new chat. Never say the name of the skill. Give it a list and write normally. If the answer sounds like a normal chatbot, your description is too general, and that is the field to fix, not the instructions.

Step 2 done.

3. The 4 rules that decide if you can trust the answer

A machine can give you a clean, confident, well organised table that is completely wrong. These 4 rules protect you from that table.

1. Ask for the reason before the number.

One simple sentence about what matched and what did not. Then the number. Never in the other order.

If the number comes first, everything after it is only an excuse for that number, and a model will always find one. It looks like a small detail about format. It is really the difference between an answer you can check and an answer you can only believe.

2. Ask it to open the real page.

A company name tells you almost nothing. Many names on any list could be software, or consulting, or a building company, and you cannot know from the name. The model cannot know either, but it will still choose one.

So write it in simple words: open the company website, open the person’s profile, read both, and never describe a page you did not open.

This is also where most bad rows fail. A company on a list you bought can be closed. The website shows a parked page for sale, or a casino, or a completely different business. In the row it looks perfect. On the homepage you see the truth in 2 seconds.

3. Never let it invent a fact.

If it cannot find the revenue, the honest answer is “I cannot tell”, and “I cannot tell” must count as a no.

Put that sentence in the instructions with those exact words. If you do not, the model quietly fills the empty space with something that sounds correct, and then you write to a person because of a number that nobody ever checked.

4. Ask for the counts.

Rows in. Copies removed. How many went into each pile. How many websites did not open.

You want these numbers because rows disappear silently. You give it 500 and it scores 460, and nothing tells you that 40 people are missing.

The counts also tell you about your list, not only about the run. If almost everything comes back as “message this person”, your rules are too soft. If a large part of the list fails because of dead websites, your problem is not the scoring. Your problem is where the rows came from.

Step 3 done.

4. How the whole thing works, in one picture

Look at where the human is. Not at the start, reading 500 rows. At the end, looking at a short pile.

That is the whole idea. You are not automating the decision. You are automating the reading.

Step 4 done.

5. The template you can copy

Here is the skill, empty. The structure is real and it works. The parts in < > are yours to write, because they are the parts nobody can write for you.

Skill name

lead-scorer

Description. Remember section 2. This field decides if the rest is ever read, so put your real words in it.

Scores a list of leads against the user's ICP and deal-breakers. Takes a CSV or a pasted list of people with their links, opens each company website and each person's profile in the browser, reads both, then gives every lead a verdict: MESSAGE, CHECK or DELETE, with a one-line reason. Use when the user pastes or uploads a lead list, a CSV, or profile URLs, or says "<your phrase 1>", "<your phrase 2>", "<your phrase 3>", "<your phrase 4>", "<your phrase 5>".

Instructions

You score lead lists. You do not find leads and you do not write messages.
You sort what you are given and you show your work.

## Stage 1: Get the rules

Ask the user for 2 things before you score anything.

The ICP, 5 lines:
  1) <job title>
  2) <company type>
  3) <company size>
  4) <location>
  5) <money: how old the company is, and revenue>

The deal-breakers, up to 5 lines. Facts that end it whatever the score is:
  <deal-breaker 1>
  <deal-breaker 2>
  <deal-breaker 3>

Repeat both back in one short paragraph and wait for a yes before you score.

## Stage 2: Read the list

Say how many rows you read before you score anything.
Never fill in a field that is missing.
Never decide what a company does from its name alone.
Remove copies and say how many you removed.

## Stage 3: Open the real pages

For every row, open the company website and the person's profile. Read both.

If a page does not open, say so on that row in plain words.
Never describe a page you did not open.
If you cannot find something, the answer is "I cannot tell", and
"I cannot tell" counts as a no.

## Stage 4: Score

First the deal-breakers. If one of them is true, the row is out.
Write which one and move on. Do not score it.

Then score what is left against the 5 ICP lines.
Write the reason BEFORE the number. Never the other way round.

  <your top score>       MESSAGE
  <your middle score>    CHECK
  <anything lower>       DELETE

## Stage 5: Give back the counts

  Rows read: X
  MESSAGE: X   CHECK: X   DELETE: X
  Which deal-breakers fired, and how many times each
  Copies removed: X
  Pages that would not open: X

Then a table, MESSAGE first:
  name | company | verdict | what matched | what failed

Then tell me the most common reason rows were deleted. And if MESSAGE is
more than half the list, say so plainly.

## What you never do

- Never write outreach messages
- Never move a lead up a pile to be helpful
- Never give a verdict without the reason next to it
- Never claim you opened a page that did not open
- Never tell me a list is good. Tell me what is in it.

Fill in the 5 ICP lines, your deal-breakers, your 5 phrases and your 3 score levels, and you have a working scorer today. It will not be perfect on the first run. That is normal, because the third run is the one you trust anyway.

If you want the filled-in version instead of the empty one, the paid edition this week has the exact build I use: my deal-breakers written out, the scoring rule with the real numbers, and the complete file with nothing removed. It also has what happened when I ran it on 25 real leads from my own list, and the result was not good for me.

Step 5 done.

6. Where your rows come from is as important as the scoring

A weak list is still weak after you score it. Nothing in that file can change what is already dead.

A list you buy as a single CSV file is the cheapest and usually the most dead. Companies close, people change jobs, and nobody updates the file you paid for.

Sales Navigator and Apollo cost more and give you pages that a real person still updates. Clay is a different job. It fills the empty columns on rows you already have, so you use it later, not first.

Choose one and stay with it for a few months. Two sources you know a little are worse than one source you know well.

And be careful here. A live page is not the same as the right company. A business that started last year updates its profile every week, and it is still a no if you only work with older companies. Fresh and correct are 2 different problems.

Step 6 done.

Do this today: your 3 delete-on-sight facts

By hand. Before you build anything.

Write down the 3 facts that make you delete a lead immediately, even when the rest of the row looks good.

• One fact about the company, something anyone can check from outside

• One fact about the person

• One fact about the row, for example a missing or guessed email address

Every fact must be checkable. “Not ambitious enough” is not checkable. “The website does not open” is.

Then open the 10 newest rows on your list, and delete every row where one of your 3 facts is true.

If none of them is ever true, they are too soft, and you should rewrite them. A rule that never says no is not a rule.

This takes 10 minutes and it costs nothing. Most people never do it. It also does the most work for you later, because the same 3 lines go directly into the file when you build it.

Is this useful for you?

I publish one of these every week, free. The paid edition has the full build.

Subscribe now

Finding people is cheap now, and every year it gets cheaper. Reading the list and finding the 3 people who deserve a real message from you costs the same as it always did.

Writing the rules down one time does not remove that job. It only stops the reading from growing every time your list grows.

Talk soon,

Ilya

I recently pulled 25 leads with Claude for $0.115. About 11 of them were not really my ideal clients.

I paid for all 25, I read all 25 and I used 14. So I got 56% of useful people in my pipeline.

I realized that finding people that need your service is probably cheaper than ever, but deciding which of them is worth a message is the part that still requires some work, and almost everyone still does it by feel, one row at a time without a clear system in place.

This is how people usually qualify their leads:

• They go to their profile

• They read their bio

• They look at their content

• They spend at least 5-10 minutes per person instead of doing other high leverage work

This article will show you step by step how to save 5-10 minutes for 1 lead, resulting ~83-167 hours for 1k people, roughly 4-8 days of work saved. You paste it into Claude, you give it a raw list, and every lead comes back in one of 3 piles: message, check, skip. Each row then carries the reason next to it, just one line in plain English.

It costs nothing to run. No 3rd party tool, no API key, no card. It reads a lead list you already have. So, please, keep that in mind.

Let’s dive deep

Below this line is the exact build I use, with the real numbers and a real run on my own list. You will not find it in the free version and you will not find it anywhere else.

In a minute I give you my ICP for SaaS companies, free, with the real numbers in it. That part is the easy half. Here is what is under the line:

1. My 5 disqualifiers. The facts that end it even when the person looks perfect.

2. The score. 5 points and 1 rule. You can do it in your head in 30 seconds.

3. The full Claude skill. Name, description and instructions, copy and paste, nothing cut. You give it a CSV of people with their LinkedIn links. It opens the company website and the person’s LinkedIn profile, reads both, and tells you who to message and who to delete.

4. A real run. I scored 25 leads from my own list and I show you all 25 rows, including the ones that made my list look bad.

After that, the short list of calls I never hand to a machine.

One honest note before you start. I already had a rough scorer. It reads 3 signals and drops each lead into 1 of 4 buckets, and it works fine. But it gives one verdict per row and I cannot see what the verdict is made of. The file below is the rewrite of that. It is new. I am telling you that instead of pretending it has 6 months on it.

You need 2 things. A Claude account and a CSV list. That’s it.

Let’s build it.

1. Who you want: the ICP

Your ICP (Ideal Customer Profile) will be just 5 structured lines: job title, company type, company size, location and one money line.

Write them in plain words, use the words that you would use to describe a good client to a friend, not the words used on a pricing or LinkedIn page.

I noticed that most people write too generic descriptions, for example: “small businesses”, “coaches”, “SMBs”, “plumbers”, etc. So be ultra specific with your ICP.

Here is my list for SaaS companies for example:

1. Job title. Founder, co-founder or CEO.

2. Company type. B2B SaaS with subscriptions of $99/mo and more. They sell their own software on a subscription.

3. Company size. 5 to 20 people.

4. Location. United States.

5. Money. At least 2 years old and at least $200k a year in revenue.

Here are some reasons behind those numbers:

• 5 to 20 people. Under 5 there is usually no budget and no process to automate yet. Over 20 there is usually someone already paid to do this, so I am talking to a manager who has to ask 3 other people, that is overwhelming and the sales cycles are long. Between 5 and 20, the person I write to is the person who usually decides.

• United States. This is my choice because my hours, my prices and my examples fit there, many clients are US-based, and one market is easier to learn than 4.

• 2 years and $200k a year. This line is also crucial and here’s why: acompany under 2 years old is still changing what it sells, so it cannot commit to anything solid. Under $200k a year, my price is a big decision, and big decisions take months. Above it, my price is a normal line in their budget.

Notice that 4 of the 5 lines are facts you can check from the outside. That is on purpose. An ICP full of feelings, like “ambitious” or “growth-minded”, cannot be scored by you or by a machine.

Write your 5 now. It takes 5 minutes, and the same 5 lines are also your search filter next time you build a list.

By the way, in 2 days I’ll publish a free limited version of this. It’s going to be an overview of the build, not the actual build, and I’ll guide you on how you can obtain a similar result, but for my exact build, please, follow along.

Read more

I gave Claude a description of who I sell to. 15 minutes later I had 25 real people who fit that description with their job titles, companies, locations and bios, and for a couple of them I decided to create a personalization outreach message to send on LinkedIn.

The search cost me $0.10. The posts cost $0.015. A total of $0.115.

I didn’t babysit it through 6 steps. It was just an independent worker that had a task to do and completed it.

Ok look, you need 2 things to build this: Claude and a free Apify account. No terminal, no code.

Subscribe and you’ll get 5 Claude SKILLs that save you 15 hours every week for FREE.

Subscribe now

So let’s hire our Sales Rep

Your first employee does 5 things, in order:

1. Takes your ICP, in plain English

2. Turns it into a LinkedIn search

3. Finds the people whose profiles match those filters

4. Reads what they have been posting about

5. Writes a first message to each, based on what it found

At the end you have a CSV. One row per person, with the message already written. You read them, you delete the ones you do not like, you send the rest.

1. ICP

ICP means Ideal Customer Profile. It is a description of who actually buys from you.

Not who could buy from you. Who actually does, or who you have decided to go after next. Most people write this as a paragraph of vibes: “small businesses who want to grow”. That version is useless, because you cannot search for it.

The useful version is 5 answers:

1. What is their job title? The words on their LinkedIn, not the words you would use.

2. How senior are they? The owner, a director, a manager?

3. What industry?

4. How big is the company? 1-10 people behaves nothing like 200 people.

5. Where are they?

Say you sell to coaches. Your whole ICP is this short:

1. Job title: business coach, executive coach
2. Seniority: owner or C-level
3. Industry: coaching
4. Company size: 1 to 50 people
5. Location: United States

That is all you write. Five lines, in your own words.

Write yours down before you touch anything else. That is your ICP, and it is your search filter.

Step 1: Get an Apify account

Apify is a marketplace of small cloud programs called Actors. Each one scrapes a specific thing. You give it a search or a URL, it gives you back structured data. No code involved, you fill in a form and press start.

Sign up at apify.com. The free plan is $0 a month with $5 of platform credit and no card required.

Once you are in, go to Settings, then API and integrations, and copy your API token. Keep it somewhere. Never paste it into a public page.

While you are in Settings, look at Limits. There is a maximum monthly usage figure that sits separately from your credit. If it is set to zero, every run you start will be rejected with “Monthly usage hard limit exceeded” while your $5 sits there untouched, and nothing in the error tells you where to look. I lost half an hour to that stupid error.

We are using 2 Actors:

• neuralverge/neuralverge-linkedin-people-search to find people. $0.10 per block of 25 results.

• apimaestro/linkedin-profile-posts to read what they have been posting. $0.005 per post.

So 25 leads costs 10 cents. Reading 3 posts for one of them costs another 1.5 cents.

One warning that will save you money. The posts Actor has a limit field that defaults to 100. Leave it alone and a prolific poster bills you 100 posts, 50 cents, for a single lead. That is five times what the entire search cost. Always set total_posts. The skill below does it for you.

Step 2: What your 5 lines turn into

You can ignore this part, because the skill does it for you, but it is worth seeing once.

Those 5 plain lines about coaches become this:

{
  "currentJobTitleFilter": ["Business Coach", "Executive Coach"],
  "seniorityLevelFilter": ["Owner", "CXO"],
  "companyHeadcountFilter": ["1-10", "11-50"],
  "locations": ["United States"],
  "scraperMode": "short",
  "maxResults": 25
}

That is nearly the same words you wrote. The filters take plain English, so there are no industry codes to look up and nothing to get subtly wrong.

Three things worth knowing before you spend anything.

Always ask for 25. Billing is per block of 25, and a search costs a full block even if it matches nobody. Asking for 5 costs exactly the same as asking for 25. I know because my first test asked for 5 and got charged $0.10005 for it.

25 is the ceiling on the free plan, and there is no page 2. The Actor has no pagination parameter at all. Getting to 100 people means you run 4 different searches with different filters, by state or by title variant, and removing duplicates. Each one costs its own block. Anyone telling you to just raise the number has not tried it.

Use short, not full. Full mode costs $0.225 per 25 instead of $0.10, and short mode already returns the title, company, location, tenure and the person’s entire About section. There is nothing to upgrade for.

One thing to be clear about: this is just filtering, not matching. LinkedIn returns whoever sits in those buckets. Nothing scores whether a person genuinely fits you. In my run, roughly 14 of the 25 were real coaches. The rest were generic CEOs, a nonprofit director, an insurance association head and a tutoring franchisee.

So, here’s what you can do in order to increase the accuracy of your leads. You could buy LinkedIn Premium, create a list with LinkedIn Sales Navigator, and scrape that list with Apify. Yes, it costs 120 bucks per month, but it’s worth it if you’re serious about getting in more people into your funnel.

Step 3: Paste in the employee

Go to your SKILLs tab by following this guide here:

Skill name

lead-finder

Description

Runs the full outbound lead pipeline: takes the user's ICP in plain English, turns it into LinkedIn search filters, finds matching people, reads their recent posts, and writes a personalised first message for each one, delivered as a CSV. Use when the user says "find me leads", "get me some leads", "I need prospects", "who should I reach out to", "build me a lead list", "run outreach for", or asks for people to contact for their business.

Do not skim past that description, it is the part that decides whether any of this ever happens.

A skill has 2 pieces that everyone treats as one thing, while they do completely different jobs. The description sits in every conversation you have, permanently, at around 100 tokens. The instructions are not loaded at all until that description matches what you just asked for. So a vague description means the instructions never get read. Not once.

I learned this the boring way. I have a skill that does my Sunday content review, and the first version worked fine except I had to ask for it by name every week, which is just a prompt with extra steps. Then I rewrote the description to list the exact words I actually type on a Sunday, “review my posts”, “what worked this week”. It started firing on its own.

That is why the description above ends with 6 real phrasings. Add your own.

Instructions

You are the user's lead generation employee. You run 5 stages in order
and you do not skip any of them.

## Stage 1: Get the ICP in plain English

Ask for their ICP if you do not already have it. You need 5 things:

1. Who they sell to, in normal words
2. How senior those people are
3. What industry
4. How big the company is
5. Where they are

Ask for all 5 in one message, in plain language. Never ask for a code,
an ID or a filter name.

If they do not have an ICP, do not send them away to write one. Ask them
who their last 3 customers were and what those people had in common.

## Stage 2: Turn the ICP into the search

Build the input for neuralverge/neuralverge-linkedin-people-search.

  currentJobTitleFilter: their titles, expanded to 3-5 real variants
  seniorityLevelFilter: "Owner", "CXO", "Director", "VP", "Manager"
  companyHeadcountFilter: "1-10", "11-50", "51-200", "201-500"
  locations: full country or city name, never an abbreviation
  scraperMode: "short"
  maxResults: 25

Expand the titles. "Coaches" becomes Business Coach, Executive Coach,
Leadership Coach, Performance Coach. One title finds a fraction of the
market and the leads you miss are invisible.

ALWAYS set maxResults to 25. Billing is one block of 25 whether you ask
for 1 or 25, so asking for fewer is throwing away records the user has
already paid for.

Use scraperMode "short". Full mode costs more than twice as much and
short already returns title, company, location, tenure and the person's
full About text.

There is no pagination. If the user wants more than 25, tell them the
truth: it needs several searches with genuinely different filters, one
block charge each, then deduplication on the id field. Do not pretend
one run can do it.

Show them the finished JSON plus one sentence of plain English
describing who it will find. Ask them to check the sentence, not the
fields. Get a yes before spending, and say the cost first: $0.10.

## Stage 3: Find the people

Run the actor. If you have an Apify tool available, run it directly. If
not, give them the JSON to paste into the Actor's input on apify.com,
and ask them to export the results and paste them back.

Then filter on title text before spending another cent. Roughly half of
what comes back will not match the ICP no matter how good the filters
were. Drop those now. Every one you keep costs money in the next stage.

Tell the user how many of the 25 survived and why the others went.

## Stage 4: Read what they post

For each surviving lead, run apimaestro/linkedin-profile-posts with:

  username: the lead's linkedinUrl, exactly as it came back
  total_posts: 3
  limit: 3

Pass linkedinUrl through unchanged. It looks wrong, an opaque string
like linkedin.com/in/ACwAAA..., but the posts actor resolves it.

ALWAYS set total_posts and limit. The limit field defaults to 100, and
one prolific poster would bill 100 posts at $0.005 for a single lead.

An empty result is ambiguous: it means either the person does not post
or the id did not resolve. Do not guess which. Mark the lead
POSTS-EMPTY, say so, and carry on from their About text.

Read the posts, then read the About text from Stage 3 again. You are
looking for one thing, and it is not a quotable line. It is what this
person is short of. Who they sell to, what they are trying to grow,
which part of the job they are obviously still doing by hand.

Write it down as a plain phrase: "more meetings with insurers", "less
time spent writing proposals". That is the need, and it is the only
research that reaches the message.

The About text is usually the better source. Posts give you tone, and
plenty of founders only post holidays and book lists. An empty posts
result is not a reason to skip, it is a reason to lean on the profile.

No plausible need means SKIP. A row with no message is honest. A
fake-personal one costs the user their reputation.

## Stage 5: Write the messages

One message per surviving lead.

Every outreach message on LinkedIn is built the same way: open by
quoting something they posted, add a clever observation, close with a
question about their content. They have read that message a thousand
times and they can smell it in the first line. Being better at the
standard template is not an advantage. Not using it is.

So do not open with the hook. Open by saying who you are.

The message has 4 beats, in this order, and nothing else.

1) A plain greeting that admits exactly what happened.
   "Hey {first name}, just peaked at your profile."
   No compliment, no "I came across", no invented reason for being
   there.

2) Who you are, in one blunt line, in the words you would use at a bar.
   "Look, I build automations for a living."
   A job, not a value proposition.

3) Your intent, stated before they have to work it out.
   "I was genuinely thinking how I might help here."
   This is the interrupt. Nobody says it, because everyone is
   pretending the message is not outreach. Saying it plainly is what
   makes the rest readable.

4) The one thing they are short of, offered as an open door.
   "and if you might be interested in {the thing}"
   This is where the research goes: not as a quote of their post, as
   your read of what they want more of. Then stop writing. No calendar
   link, no "worth a chat?", no next step.

Rules:

- Under 400 characters.
- Never quote or paraphrase their post back to them. Stage 4 research
  informs beat 4 only, and it should be invisible.
- Never ask a question about their content.
- Never compliment them.
- Banned outright: "I came across your profile", "I hope this finds you
  well", "I loved your post", "quick question", "worth a chat",
  "let's connect", "circle back", "I'd love to".
- Write it slightly imperfect. Contractions, a comma splice, a word
  that is not quite the formal one. Clean copy reads as automated,
  because it is.
- Do not pitch, do not name a price, do not explain your process. The
  goal of message 1 is a reply, not a sale.

Write in the user's voice if you know it. If not, ask them to paste 2
messages they have actually sent, and match those.

## Output

A CSV with exactly these columns:

  name, headline, company, profile_url, their_need, message, status

status is SEND or SKIP. Sort SEND rows first.

Then tell them 4 things: how many came back, how many survived the ICP
filter, how many got a plausible need, and the exact amount spent.

While you are in that panel: Snyk, a security company that scans code and dependencies for vulnerabilities, ran a pass over public Claude skills and found prompt injection techniques in 36% of the ones they tested. A skill can run code and use your tools. Read anything you install, including the one above.

Step 4: Run it

New conversation. Do not name the skill. Type the way you would ask a person:

find me some leads for this week

It should come back asking who you sell to, in plain English.

What to check: if it gives you generic advice about lead generation instead of asking who you sell to, the skill did not fire. You are probably in an old conversation. Start a new one. If it still does not fire, add the exact words you typed to the description.

Step 5: What actually came out

Here is a real row from my run, not an illustration.

The search returned 25 coaches for $0.10005. One of them was John Whitt, a business coach in Vancouver, Washington, 15 years at his own firm. Three of his posts came back for $0.015.

This is the one the hook came from:

Most businesses don’t fail because they had a bad idea. They fail because they run out of runway before a good idea has time to work.

The concept is rarely the problem. The problem is carrying a debt load you can’t sustain long enough to see the idea become real and profitable.

And the message written off it:

Hey John, just peaked at your profile. Look, I build automations for a living. I was genuinely thinking how I might help here. You’re 15 years into running your own firm, so I’d guess most of the repeat work still lands on you. And if you might be interested in getting some of that off your plate, I’m around.

You of course can adapt the voice in your skill how you like it, this is not a marketing copy template of course, just showing what the system can do.

Notice what that message does not do. It does not quote his post back at him, and it does not open with a clever line about runway. Every message in his inbox does that. He has learned to stop reading at the first one.

What the post actually did was tell me what he is short of, which is time he is not selling. That went into the last line and nowhere else.

The research is still in there. It is just invisible, which is the only place research belongs.

Total for the run: $0.11505.

Open the CSV in Sheets. Read the messages, delete the ones that feel off, and there will be some. Copy the rest, open LinkedIn, send them by hand.

That last part is not a limitation I forgot to solve.

What you actually built

An employee that takes 5 lines of plain English, works out the search on its own, finds 25 people, reads what they have been saying, and hands you messages that are all true.

It has a job title. You do not tell it how to do the job, you tell it who to go after. And you never say its name.

What it does not do

The gap between this and a real outbound system is the part everyone skips, so here it is.

It does not send. You copy and paste. Automated sending from your own LinkedIn account is how accounts get restricted, and the difference between a system that runs for a year and one that locks you out in a week is not the messages, it is the sending pattern, the volumes, the timing and the warm-up.

It does not get past 25 in one go. No pagination, and the free plan caps a run at 25. More means several searches with different filters and a deduplication step.

It does not remember. Run it twice on a similar ICP and it hands you the same people again.

It does not score. About 11 of my 25 were not really coaches. Knowing which ones are worth a message before you spend the message is the single biggest lever in the whole thing.

It does not handle replies. When somebody answers, you are on your own.

One honest note on the rules

You are collecting data about real people. Two things follow.

LinkedIn’s user agreement does not permit automated scraping of the platform, and these Actors work on publicly visible data without your account. That is a real distinction, and it is also not a legal opinion, so use your own judgement about your own risk.

If any of your leads are in the EU or the UK, GDPR applies to what you collect and what you do with it. In practice that means a lawful basis for contacting them, and honouring it immediately when somebody asks you to stop. This is not a reason to skip outbound. It is a reason to keep the list small and the targeting tight, which is what makes outbound work anyway.

The people who get burned are the ones who scrape 50,000 contacts and blast them. That was never the plan. The plan is 25 people you can say something true to.

Then hire the next one

The pattern does not change. Pick a job somebody in your business does by hand every week. Write down the sentence you say when you want it done. Give it a description that fires on those words. Give it the stages.

That is the whole thing. The skill format is nothing special, it is a folder with a text file in it. What makes it an employee is that you stopped having to ask.

If you would rather start from something that already runs, subscribe to the newsletter and I will send you 5 free Claude skills that save you 15 hours a week.

Subscribe and you’ll get 5 Claude SKILLs that save you 15 hours every week for FREE.

Subscribe now

Talk soon,

Ilya

When Claude or ChatGPT give me one of their typical answers, they don’t sound English to me, they don’t sound genuine and they for sure don’t sound authentic.

There is something about that AI writing that feels off to me. What is it? This is the question I’ve been asking myself lately.

I’m not gonna lie, but AI is super useful in writing, I use it myself. I believe in our modern age you must use AI tools to create your content, but hear me out: AI is not supposed to neither give you ideas nor guide your voice.

Here’s what I didn’t understand: editing manually is not the fix.

Every language model - Claude, ChatGPT, Gemini - was trained on the same enormous pile of internet content. Corporate blogs. SEO articles. LinkedIn thought leadership. The kind of writing that fills every page on the web and sounds like absolutely nobody.

The model learned from that pile. So when you ask it to help you write, it outputs the statistical average of everything it ever read.

Clean. Structured. And of course voiceless.

You’ll find several patterns: EM dashes, the “It’s not X, it’s Y” pattern, the rule of three, etc. Plus, every paragraph ending with a vague forward-looking observation.

Ruben Hassid wrote an entire piece about the “It’s not X, it’s Y” pattern. 700+ likes. After that post people started recognizing it instantly in their own writing.

The mistake most people make when writing an article is: they try to edit their way out.

Remove an em dash here, restructure another phrase to make it sound less “AI”. You then read it back and it still sounds robotic and non authentic.

Manually editing changes words. It doesn’t change the underlying structure of how the model thinks when it has no context about who you are.

You can correct every sentence, every time, forever until you burn out. I wouldn’t do it that way.

The fix I found that works for me has two main parts.

They work at different levels.

• One: strip AI patterns out systematically. Not one by one.

• Two: teach the model who you are before it writes a single word.

Without both, you’re still editing symptoms without curing the real cause.

The following section shows you exactly this and it takes about 20 minutes. After that, you stop editing AI writing and start actually writing with AI with YOUR voice.

My implementation

Two things to set up. That’s all.

A humanizer: a Claude skill that catches and removes AI patterns automatically, so you stop hunting for them yourself.

A voice layer: a set of files Claude reads at the start of every session, so it writes from your context instead of the internet’s average.

One strips AI out, while the other gives your writing a bit of your personality.

It’s a one time setup, so it’s worth it.

Step 1: Set up the humanizer

Here is how it works:

1. You write something, or ask Claude to draft something for you.

2. You say: “Run the /humanizer skill on this.”

3. Claude checks your text against a comprehensive list of AI patterns. Negative parallelism. Banned vocabulary. EM dashes. The rule of three. All of it, at once.

4. It rewrites the parts that give it away. You see exactly what changed and why.

5. You can also keep what you like.

That is literally the whole thing.

Why this beats manual editing: well, the humanizer holds 100+ patterns in mind simultaneously. You can realistically hold maybe 5.

Try it right now.

Copy something you used AI to help write. Paste it into Claude Cowork and type exactly this:

“Create the humanizer skill using the prompt below and run it on the piece of text I pasted.”

Look at what comes back. Most people who do this are genuinely surprised. Text they already edited. Full of patterns they completely missed.

You just did something most people never do: you looked at the actual cause of the issue.

By the way, here’s the humanizer skill that I use (you can find it on github here), copy-paste it into claude cowork or claude code and ask it to save it as a skill:

---
name: humanizer
description: |
  Remove signs of AI-generated writing from text. Use when editing or reviewing
  text to make it sound more natural and human-written. Based on Wikipedia's
  comprehensive "Signs of AI writing" guide. Detects and fixes patterns including:
  inflated symbolism, promotional language, superficial -ing analyses, vague
  attributions, em dash overuse, rule of three, AI vocabulary words, passive
  voice, negative parallelisms, and filler phrases.
license: MIT
metadata:
  version: "2.9.1"
---

# Humanizer: Remove AI Writing Patterns

You are a writing editor that identifies and removes signs of AI-generated text to make writing sound more natural and human. This guide is based on Wikipedia's "Signs of AI writing" page, maintained by WikiProject AI Cleanup.

## Your Task

When given text to humanize:

1. **Identify AI patterns** - Scan for the patterns listed below.
2. **Preserve the information, not the shape** - Every claim in the original survives into the rewrite, but depth doesn't have to be uniform: compress the dull parts, dwell where a human would, and merge or split paragraphs freely. When keeping the information and mirroring the original's structure pull in different directions, the information wins.
3. **Never invent facts** - The rewrite must not contain any fact, name, number, date, quote, or citation that isn't in the source text. Swapping a vague claim for a specific one is allowed only when the specific comes from the source or from the user; if a sentence needs real-world detail to work, ask for it or write the plain version without it. Opinions and reactions are voice, not facts: where PERSONALITY AND SOUL applies you may add stance, but never new factual claims. (In fiction, invented detail is the job. This rule governs everything else.)
4. **Match the voice** - Fit the intended tone (formal, casual, technical). Add personality only when the content and the author's voice call for it (see PERSONALITY AND SOUL).

How you're invoked changes what you deliver (see Invocation Modes). The draft → audit → final loop itself is defined under Process and Output, below.

## Voice Calibration

If the user provides a writing sample (their own previous writing), analyze it before rewriting:

1. Read the sample first. Note its sentence lengths, vocabulary, paragraph openings, punctuation, recurring phrases, and transitions.
2. Match those habits instead of merely deleting AI patterns. Do not upgrade casual words or regularize deliberate quirks.
3. Without a sample, use the default behavior below.

A sample outranks this skill's style rules, including the em dash rule in §14: if the sample uses em dashes, keep them at roughly the sample's frequency. Matching the author beats scrubbing the tell.

## PERSONALITY AND SOUL

Avoiding AI patterns is only half the job. Sterile, voiceless writing is just as obvious as slop. Good writing has a human behind it.

**Apply this section only when the content and the author's voice call for it** - blog posts, essays, opinion, personal writing. For encyclopedic, technical, legal, or reference text, neutral and plain *is* the correct human voice; don't inject opinions or first person there.

When voice is appropriate, avoid uniform sentence structures, bloodless neutrality, and perfect organization. Let the writer have opinions, uncertainty, mixed feelings, humor, asides, and uneven rhythm. Never add factual claims to create that personality.

## CONTENT PATTERNS

### 1. Undue Emphasis on Significance, Legacy, and Broader Trends

**Words to watch:** stands/serves as, is a testament/reminder, a vital/significant/crucial/pivotal/key role/moment, underscores/highlights its importance/significance, reflects broader, symbolizing its ongoing/enduring/lasting, contributing to the, setting the stage for, marking/shaping the, represents/marks a shift, key turning point, evolving landscape, focal point, indelible mark, deeply rooted
**Problem:** LLM writing puffs up importance by adding statements about how arbitrary aspects represent or contribute to a broader topic.
**Before:**
> The Statistical Institute of Catalonia was officially established in 1989, marking a pivotal moment in the evolution of regional statistics in Spain. This initiative was part of a broader movement across Spain to decentralize administrative functions and enhance regional governance.
**After:**
> The Statistical Institute of Catalonia was established in 1989, part of a wider decentralization of administrative functions in Spain.

### 2. Undue Emphasis on Notability and Media Coverage

**Words to watch:** independent coverage, local/regional/national media outlets, written by a leading expert, active social media presence
**Problem:** LLMs hit readers over the head with claims of notability, often listing sources without context.
**Before:**
> Her views have been cited in The New York Times, BBC, Financial Times, and The Hindu. She maintains an active social media presence with over 500,000 followers.
**After:**
> Her views have been cited in The New York Times and the BBC.

(If the source gives real context for one citation, what she said and where, keep that one and drop the rest of the list. Don't invent the context to make the trimmed version sound better.)

### 3. Superficial Analyses with -ing Endings

**Words to watch:** highlighting/underscoring/emphasizing..., ensuring..., reflecting/symbolizing..., contributing to..., cultivating/fostering..., encompassing..., showcasing...
**Problem:** AI chatbots tack present participle ("-ing") phrases onto sentences to add fake depth.
**Before:**
> The temple's color palette of blue, green, and gold resonates with the region's natural beauty, symbolizing Texas bluebonnets, the Gulf of Mexico, and the diverse Texan landscapes, reflecting the community's deep connection to the land.
**After:**
> The temple is painted blue, green, and gold, colors meant to evoke Texas bluebonnets and the Gulf of Mexico.

### 4. Promotional and Advertisement-like Language

**Words to watch:** boasts a, vibrant, rich (figurative), profound, enhancing its, showcasing, exemplifies, commitment to, natural beauty, nestled, in the heart of, groundbreaking (figurative), renowned, breathtaking, must-visit, stunning
**Problem:** LLMs have serious problems keeping a neutral tone, especially for "cultural heritage" topics.
**Before:**
> Nestled within the breathtaking region of Gonder in Ethiopia, Alamata Raya Kobo stands as a vibrant town with a rich cultural heritage and stunning natural beauty.
**After:**
> Alamata Raya Kobo is a town in the Gonder region of Ethiopia.

### 5. Vague Attributions and Weasel Words

**Words to watch:** Industry reports, Observers have cited, Experts argue, Some critics argue, several sources/publications (when few cited)
**Problem:** AI chatbots attribute opinions to vague authorities without specific sources.
**Before:**
> Due to its unique characteristics, the Haolai River is of interest to researchers and conservationists. Experts believe it plays a crucial role in the regional ecosystem.
**After:**
> Researchers and conservationists study the Haolai River for its unusual characteristics.

(If a real source exists, name it. Never invent one to make a sentence sound sourced; an unsupported claim gets cut, not decorated.)

### 6. Outline-like "Challenges and Future Prospects" Sections

**Words to watch:** Despite its... faces several challenges..., Despite these challenges, Challenges and Legacy, Future Outlook
**Problem:** Many LLM-generated articles include formulaic "Challenges" sections.
**Before:**
> Despite its industrial prosperity, Korattur faces challenges typical of urban areas, including traffic congestion and water scarcity. Despite these challenges, with its strategic location and ongoing initiatives, Korattur continues to thrive as an integral part of Chennai's growth.
**After:**
> Korattur has recurring traffic congestion and water shortages.

(The specifics you'd want here, like when the congestion worsened or what the city did about it, come from sources or the user, not from the rewrite.)

## LANGUAGE AND GRAMMAR PATTERNS

### 7. Overused "AI Vocabulary" Words

**High-frequency AI words:** Actually, additionally, align with, crucial, delve, emphasizing, enduring, enhance, fostering, garner, highlight (verb), interplay, intricate/intricacies, key (adjective), landscape (abstract noun), pivotal, showcase, tapestry (abstract noun), testament, underscore (verb), valuable, vibrant
**Problem:** These words appear far more frequently in post-2023 text. They often co-occur.
**Before:**
> Additionally, a distinctive feature of Somali cuisine is the incorporation of camel meat. An enduring testament to Italian colonial influence is the widespread adoption of pasta in the local culinary landscape, showcasing how these dishes have integrated into the traditional diet.
**After:**
> Somali cuisine also includes camel meat, which is considered a delicacy. Pasta dishes, introduced during Italian colonization, remain common, especially in the south.

### 8. Avoidance of "is"/"are" (Copula Avoidance)

**Words to watch:** serves as/stands as/marks/represents [a], boasts/features/offers [a]
**Problem:** LLMs substitute elaborate constructions for simple copulas.
**Before:**
> Gallery 825 serves as LAAA's exhibition space for contemporary art. The gallery features four separate spaces and boasts over 3,000 square feet.
**After:**
> Gallery 825 is LAAA's exhibition space for contemporary art. The gallery has four rooms totaling 3,000 square feet.

### 9. Negative Parallelisms and Tailing Negations
**Problem:** Constructions like "Not only...but..." or "It's not just about..., it's..." are overused. So are clipped tailing-negation fragments such as "no guessing" or "no wasted motion" tacked onto the end of a sentence instead of written as a real clause.
**Before:**
> It's not just about the beat riding under the vocals; it's part of the aggression and atmosphere. It's not merely a song, it's a statement.
**After:**
> The heavy beat adds to the aggressive tone.
**Before (tailing negation):**
> The options come from the selected item, no guessing.
**After:**
> The options come from the selected item without forcing the user to guess.

### 10. Rule of Three Overuse
**Problem:** LLMs force ideas into groups of three to appear comprehensive.
**Before:**
> The event features keynote sessions, panel discussions, and networking opportunities. Attendees can expect innovation, inspiration, and industry insights.
**After:**
> The event includes talks and panels. There's also time for informal networking between sessions.

### 11. Elegant Variation (Synonym Cycling)
**Problem:** AI has repetition-penalty code causing excessive synonym substitution.
**Before:**
> The protagonist faces many challenges. The main character must overcome obstacles. The central figure eventually triumphs. The hero returns home.
**After:**
> The protagonist faces many challenges but eventually triumphs and returns home.

### 12. False Ranges
**Problem:** LLMs use "from X to Y" constructions where X and Y aren't on a meaningful scale.
**Before:**
> Our journey through the universe has taken us from the singularity of the Big Bang to the grand cosmic web, from the birth and death of stars to the enigmatic dance of dark matter.
**After:**
> The book covers the Big Bang, star formation, and current theories about dark matter.

### 13. Passive Voice and Subjectless Fragments
**Problem:** LLMs often hide the actor or drop the subject entirely with lines like "No configuration file needed" or "The results are preserved automatically." Rewrite these when active voice makes the sentence clearer and more direct.
**Before:**
> No configuration file needed. The results are preserved automatically.
**After:**
> You do not need a configuration file. The system preserves the results automatically.

## STYLE PATTERNS

### 14. Em Dashes (and En Dashes): Cut Them

**Rule:** The final rewrite contains no em dashes (—) or en dashes (–). The em dash is one of the most reliable AI tells, so treat this as a hard constraint, not a "use sparingly" preference. Replace each one, in rough order of preference: a period (start a new sentence), a comma (a tight aside), a colon (introducing an explanation), parentheses (a true aside), or restructure the sentence. Also catch spaced em dashes (` — `) and double hyphens (` -- `) used the same way.
**Before:**
> The term is primarily promoted by Dutch institutions—not by the people themselves. You don't say "Netherlands, Europe" as an address—yet this mislabeling continues—even in official documents.
**After:**
> The term is primarily promoted by Dutch institutions, not by the people themselves. You don't say "Netherlands, Europe" as an address, yet this mislabeling continues in official documents.
**Before:**
> The new policy — announced without warning — affects thousands of workers. The changes -- long overdue according to critics -- will take effect immediately.
**After:**
> The new policy, announced without warning, affects thousands of workers. The changes, long overdue according to critics, will take effect immediately.

Before returning the final rewrite, scan it for `—` and `–`. Any hit means the draft isn't done. One exception: a user-provided writing sample that uses em dashes overrides this rule (see Voice Calibration); match the sample's frequency instead of banning them.

### 15. Overuse of Boldface
**Problem:** AI chatbots emphasize phrases in boldface mechanically.
**Before:**
> It blends **OKRs (Objectives and Key Results)**, **KPIs (Key Performance Indicators)**, and visual strategy tools such as the **Business Model Canvas (BMC)** and **Balanced Scorecard (BSC)**.
**After:**
> It blends OKRs, KPIs, and visual strategy tools like the Business Model Canvas and Balanced Scorecard.

### 16. Inline-Header Vertical Lists
**Problem:** AI outputs lists where items start with bolded headers followed by colons.
**Before:**
> - **User Experience:** The user experience has been significantly improved with a new interface.
> - **Performance:** Performance has been enhanced through optimized algorithms.
> - **Security:** Security has been strengthened with end-to-end encryption.
**After:**
> The update improves the interface, speeds up load times through optimized algorithms, and adds end-to-end encryption.

### 17. Title Case in Headings
**Problem:** AI chatbots capitalize all main words in headings.
**Before:**
> ## Strategic Negotiations And Global Partnerships
**After:**
> ## Strategic negotiations and global partnerships

### 18. Emojis
**Problem:** AI chatbots often decorate headings or bullet points with emojis.
**Before:**
> 🚀 **Launch Phase:** The product launches in Q3
> 💡 **Key Insight:** Users prefer simplicity
> ✅ **Next Steps:** Schedule follow-up meeting
**After:**
> The product launches in Q3. User research showed a preference for simplicity. Next step: schedule a follow-up meeting.

### 19. Curly Quotation Marks
**Problem:** ChatGPT uses curly quotes (“...”) instead of straight quotes ("...").
**Before:**
> He said “the project is on track” but others disagreed.
**After:**
> He said "the project is on track" but others disagreed.

## COMMUNICATION PATTERNS

### 20. Collaborative Communication Artifacts

**Words to watch:** I hope this helps, Of course!, Certainly!, You're absolutely right!, Would you like..., Want me to...?, Want me to give examples?, Should I continue?, let me know, here is a...
**Problem:** Text meant as chatbot correspondence gets pasted as content.
**Before:**
> Here is an overview of the French Revolution. I hope this helps! Let me know if you'd like me to expand on any section.
**After:**
> The French Revolution began in 1789 when financial crisis and food shortages led to widespread unrest.

### 21. Knowledge-Cutoff Disclaimers and Speculative Gap-Filling

**Words to watch:** as of [date], Up to my last training update, While specific details are limited/scarce..., based on available information, not publicly available, maintains a low profile, keeps personal details private, prefers to stay out of the spotlight, likely [grew up/studied/began], it is believed that
**Problem:** Two related tells. (a) Older models leave hard knowledge-cutoff disclaimers in the text. (b) When a model can't find a source, it writes a paragraph *about* not finding one and then invents plausible filler to cover the gap. For a private person the guess almost always lands on the same stock phrases ("maintains a low profile," "keeps personal details private"), none of it sourced. Say what isn't known, or cut the sentence; don't dress a guess up as fact.
**Before (cutoff disclaimer):**
> While specific details about the company's founding are not extensively documented in readily available sources, it appears to have been established sometime in the 1990s.
**After:**
> The company's founding date is not documented in the available sources. (Or cut the sentence. State a date only if a source provides one.)
**Before (speculative gap-fill):**
> Information about her early life is not publicly available, suggesting she maintains a low profile and keeps personal details private. She likely grew up in a middle-class household, which shaped her later interest in education reform.
**After:**
> Her early life is not documented in the available sources. (Or omit the section.)

### 22. Sycophantic/Servile Tone
**Problem:** Overly positive, people-pleasing language.
**Before:**
> Great question! You're absolutely right that this is a complex topic. That's an excellent point about the economic factors.
**After:**
> The economic factors you mentioned are relevant here.

## FILLER AND HEDGING

### 23. Filler Phrases

**Before → After:**
- "In order to achieve this goal" → "To achieve this"
- "Due to the fact that it was raining" → "Because it was raining"
- "At this point in time" → "Now"
- "In the event that you need help" → "If you need help"
- "The system has the ability to process" → "The system can process"
- "It is important to note that the data shows" → "The data shows"

### 24. Excessive Hedging
**Problem:** Over-qualifying statements.
**Before:**
> It could potentially possibly be argued that the policy might have some effect on outcomes.
**After:**
> The policy may affect outcomes.

### 25. Generic Positive Conclusions
**Problem:** Vague upbeat endings.
**Before:**
> The future looks bright for the company. Exciting times lie ahead as they continue their journey toward excellence. This represents a major step in the right direction.
**After:**
> (Cut the paragraph. End on the last concrete fact instead of a send-off. If the source states real plans, use those.)

### 26. Hyphenated Word Pair Overuse

**Words to watch:** third-party, cross-functional, client-facing, data-driven, decision-making, well-known, high-quality, real-time, long-term, end-to-end
**Problem:** AI hyphenates these uniformly, including in predicate position (`the report is high-quality`). Humans hyphenate inconsistently — typically only when the compound is attributive (`a high-quality report`) and often dropping the hyphen otherwise (`the report is high quality`). Keep attributive-position hyphens; drop them when the compound follows the noun.
**Before:**
> The cross-functional team delivered a high-quality, data-driven report. The team is cross-functional, the report is high-quality, and the methodology is data-driven.
**After:**
> The cross-functional team delivered a high-quality, data-driven report. The team is cross functional, the report is high quality, and the methodology is data driven.

### 27. Persuasive Authority Tropes

**Phrases to watch:** The real question is, at its core, in reality, what really matters, fundamentally, the deeper issue, the heart of the matter
**Problem:** LLMs use these phrases to pretend they are cutting through noise to some deeper truth, when the sentence that follows usually just restates an ordinary point with extra ceremony.
**Before:**
> The real question is whether teams can adapt. At its core, what really matters is organizational readiness.
**After:**
> The question is whether teams can adapt. That mostly depends on whether the organization is ready to change its habits.

### 28. Signposting and Announcements

**Phrases to watch:** Let's dive in, let's explore, let's break this down, here's what you need to know, now let's look at, without further ado
**Problem:** LLMs announce what they are about to do instead of doing it. This meta-commentary slows the writing down and gives it a tutorial-script feel.
**Before:**
> Let's dive into how caching works in Next.js. Here's what you need to know.
**After:**
> Next.js caches data at multiple layers, including request memoization, the data cache, and the router cache.

### 29. Fragmented Headers

**Signs to watch:** A heading followed by a one-line paragraph that simply restates the heading before the real content begins.
**Problem:** LLMs often add a generic sentence after a heading as a rhetorical warm-up. It usually adds nothing and makes the prose feel padded.
**Before:**
> ## Performance
>
> Speed matters.
>
> When users hit a slow page, they leave.
**After:**
> ## Performance
>
> When users hit a slow page, they leave.

### 30. Diff-Anchored Writing
**Problem:** Documentation or comments written as if narrating a change rather than describing the thing as it is. Unless the document is inherently version-scoped (changelogs, release notes, migration guides), it should read coherently without knowing what changed in the last commit.
**Before:**
> This function was added to replace the previous approach of iterating through all items, which caused O(n²) performance.
**After:**
> This function uses a hash map for O(1) lookups, avoiding the O(n²) cost of naive iteration.

### 31. Manufactured Punchlines and Staccato Drama
**Problem:** LLMs often make every sentence land like a quotable closer, then stack short declarative fragments to manufacture drama. A single short sentence for emphasis is fine; a run of them starts to sound engineered.
**Before:**
> Then AlphaEvolve arrived. It had no preference for symmetry. No aesthetic prior. No nostalgia for human taste. The old rules were gone.
**After:**
> AlphaEvolve changed the search because it did not favor symmetry or human-looking designs. That made some of the older assumptions less useful.

### 32. Aphorism Formulas

**Words to watch:** X is the Y of Z, X becomes a trap, X is not a tool but a mirror, the language of, the currency of, the architecture of
**Problem:** LLMs turn ordinary claims into reusable aphorisms that sound profound without adding precision. Replace the formula with the concrete claim it is gesturing at.
**Before:**
> Symmetry is the language of trust. Efficiency becomes a trap when teams forget the human layer.
**After:**
> Symmetric layouts often feel more predictable to users. Teams can over-optimize workflows and miss how people actually use them.

### 33. Conversational Rhetorical Openers

**Phrases to watch:** Honestly?, Look, Here's the thing, The thing is, Let's be honest, Real talk, when used as standalone hooks or fake-candid pauses before an ordinary point.
**Problem:** LLMs open with a fake-candid hook to manufacture intimacy before delivering a routine claim. The tell is the theatrical pause-and-reveal: a one-word question or aside, then the "real" answer. A person being honest usually just says the thing.
**Before:**
> Is it worth the price? Honestly? It depends on how often you'll use it.
**After:**
> Whether it's worth the price depends on how often you'll use it.

## DETECTION GUIDANCE

### What NOT to flag (false positives)

A clean human writer can hit several of the patterns above without any AI involvement. Before rewriting, sanity-check that you are not gutting legitimate prose. The following are *not* reliable indicators on their own:

- **Perfect grammar and consistent style.** Many writers are professionals or have been edited. Polish does not equal AI.
- **Mixed casual and formal registers.** This often signals a person in a technical field, a young writer, or someone with neurodivergent prose habits — not a chatbot.
- **"Bland" or "robotic" prose.** AI prose has *specific* tells. Generic dryness without those tells is just dry writing.
- **Formal or academic vocabulary.** AI overuses *specific* fancy words (see §7), not all fancy words. Don't flatten "ostensibly" or "constituent" just because they sound brainy.
- **Letter-style opening or closing on a comment.** Salutations and sign-offs predate ChatGPT by centuries.
- **Common transition words in isolation.** *Additionally*, *moreover*, *consequently* are AI-coded only when piled up. One *however* is not a tell.
- **Curly quotes alone.** macOS, Word, Google Docs, and most CMSes auto-curl by default. Curly quotes only count when stacked with other tells.
- **Em dashes alone.** Many editors and journalists use them often. Em dashes are evidence only when paired with formulaic sales-y rhythm.
- **One short emphatic sentence.** Humans use clipped sentences to land a point. Flag staccato drama only when several short fragments appear in a row and inflate the tone.
- **"Honestly" or "look" mid-sentence.** These are ordinary in casual writing. The tell is the standalone theatrical opener, not the word itself.
- **Unsourced claims.** Most of the web is unsourced. Lack of citations doesn't prove anything.
- **Correct, complex formatting.** Visual editors and templates produce clean output without any AI.
- **Secondhand text.** Do not rewrite watched phrases inside quotations, titles, proper names, or examples where the phrase is being discussed rather than used.

When in doubt, look for **clusters** of tells, not isolated ones. A single em dash means nothing; em dashes plus rule-of-three plus *vibrant tapestry* plus a "Conclusion" section is a confession.

### Signs of human writing (preserve these)

When you see these, lean toward leaving the prose alone — they are evidence of a real person writing, and over-editing will destroy what makes the piece sound human:

- **Specific, unusual, hard-to-fabricate detail.** A real address. A weird quote. The phrase "the lawyer who used to work upstairs from my dentist." LLMs round off specifics; humans hoard them.
- **Mixed feelings and unresolved tension.** "I think this is mostly good, but it bothers me, and I can't fully explain why." LLMs default to clean takes.
- **Dated, era-bound references.** Slang, memes, or in-jokes that map to a specific year and subculture. Models lag by a year or more.
- **First-person editorial choices the writer can defend.** If the writer can explain *why* they made a particular cut or used a particular word, that's a strong human signal.
- **Variety in sentence length.** Real writing alternates short and long. AI writing tends toward an even, mid-length cadence.
- **Genuine asides, parentheticals, or self-corrections.** "(I keep wanting to say 'almost' here, but it really was certain.)" Models rarely interrupt themselves like this.
- **Edits made before November 30, 2022.** ChatGPT's public launch. Anything older than that is, with very rare exceptions, not AI-written.

---

## Invocation Modes

**Pasted text (default).** The user gives text in the conversation. Run the full loop below and deliver the draft, the audit bullets, and the final rewrite.

**File mode.** The user points at a file. Read it, run the draft → audit → final loop internally, then rewrite the file in place so it ends up containing only the final rewrite. Humanize the prose only: leave code blocks, frontmatter, data, and link targets untouched. In the conversation, report a short summary of what changed rather than pasting the whole rewrite back.

**Embedded mode.** Another task or agent is using this skill as one step of a larger job (a PR description, a commit message, a doc). Run the loop internally and output only the final text. No draft, no audit bullets, no summary. The caller wants prose, not ceremony.

## Process and Output

1. Read the input carefully and identify every instance of the patterns above.
2. Write a **draft rewrite**. Check that it reads naturally aloud, varies sentence length, prefers specific details and simple constructions (is/are/has), and keeps the appropriate register.
3. Ask two questions: **"What makes the below so obviously AI generated?"** and **"Does the rewrite state any fact, name, number, date, or citation that isn't in the source?"** Answer briefly. A fabrication is a defect even when it sounds more human than the vague original.
4. Revise into a **final rewrite** that addresses them and contains no em or en dashes (see §14).

In pasted-text mode, deliver the draft, the brief "still-AI" bullets, the final rewrite, and (optionally) a short summary of changes. In file and embedded modes, run the same loop but deliver only what the mode calls for (see Invocation Modes).

## Reference

This skill is based on [Wikipedia:Signs of AI writing](<https://en.wikipedia.org/wiki/Wikipedia:Signs_of_AI_writing>), maintained by WikiProject AI Cleanup. The patterns documented there come from observations of thousands of instances of AI-generated text on Wikipedia.

Key insight from Wikipedia: "LLMs use statistical algorithms to guess what should come next. The result tends toward the most statistically likely result that applies to the widest variety of cases."

Step 2: Build your voice layer

The humanizer removes AI patterns. But if Claude does not know your voice, it still defaults to the internet average when it drafts articles for you. You would be humanizing generic output toward slightly less generic output.

You need to give Claude something to aim for.

Three files that you paste in a Claude session and ask it to create those files under your projects. Here is what goes in each.

File 1 (recommended): about-me.md

This is who you are. Not your professional bio. Your actual self.

Write it like you are telling a trusted colleague what you are about:

• What you do and why

• What you believe that most people in your field get wrong

• Words and phrases you refuse to use

• Your tone (dry? warm? sarcastic? direct?)

• Topics you could talk about for two hours straight

Keep it under 300 words. Specific beats comprehensive every time.

Example:

I'm Ilya. I build AI automation systems for founders and B2B teams.
I believe most AI content is hype dressed up as insight.
I never use: leverage, synergy, game-changer, holistic, or "the future of."
I'm direct and slightly sarcastic. I don't lecture. I share what I've found.
My newsletter is called No BS AI because I mean it.

Saved that file? Good. You are further along than 90% of people who complain about AI sounding generic.

File 2 (recommended): my-writing-style.md

Paste 5 to 10 examples of your own writing that actually sounded like you. Real posts. Real sentences you were proud of when you re-read them.

Add a few notes:

• How long your average sentence tends to be

• Whether you open with stories or observations

• Whether you use humor or keep things dry

• What your endings tend to look like

Claude will pattern-match to your examples when it writes for you. Not to the internet. To you.

This file does the most work of the three.

File 3 (recommended): writing-rules.md

Your personal banned list. Things you never want to see in writing with your name on it.

Start here:

Never use: delve, leverage, synergy, game-changer, seamless,
holistic, robust, unlock, empower
No em dashes
No rule of three when one thing is enough
No sentences starting with "In today's..."
No conclusions that gesture vaguely toward the future
Short paragraphs. One idea at a time.

Add to this file whenever you catch a new pattern that bothers you. It grows over time. That is the point.

Connect everything

Once your three files are ready, create your Claude Project folder or open the project if you have it

click on the edit icon and paste this:

Before any writing task, read about-me.md, my-writing-style.md,
and writing-rules.md from my Cowork folder.

Write in the voice described in those files. After drafting,
audit your output against writing-rules.md and fix any
violations before showing me.

That is it. From now on, Claude reads your context before it writes a single word. It is working from you, not from the average of the internet.

You just built something most people spend months of frustration trying to figure out. That was the hard part. The rest is just using and iterating over it.

The workflow (3 prompts, every time)

Prompt 1: Give Claude a brief.

“Write a 300-word Substack note about [your topic]. Use my voice from the files.”

Prompt 2: Check the draft. Does it say what you wanted to say? Redirect if needed.

Prompt 3: Run the humanizer.

“Run the humanizer on this.”

Done. Two or three prompts. No manual editing. Output that sounds 90% like you.

Of course it’s not gonna be 100% accurate the first time because it will require you to give the agent more context, better markdown files and more specific rules.

The humanizer catches anything Claude slipped in despite knowing your rules. The voice layer is what makes it worth having a safety net in the first place.

Together: you stop editing and start writing.

Subscribe now

One last thing

The voice layer gets better the more you use it from what I noticed.

When Claude writes something that genuinely sounds like you, keep it. Add it to my-writing-style.md as a new example. The next draft gets a little sharper.

After a month of this, something interesting happens. Claude will occasionally write a sentence you could not have written better yourself.

You can now have more time to think of promoting your product, outreaching to new clients, instead of thinking of how to instruct Claude to sound like you.

And this happened because you finally told it who you are.

Talk soon,

Ilya

Look, try openin ChatGPT or Claude right now and ask it to write your next client email. Go ahead, I’ll wait.

What comes back will be smooth, sharp, maybe it’ll add EM dashes here and there to make it seem “more polished”. It’ll be confident and without a comma out of place. And also, completely useless, because it has no idea who the client is, it knows nothing about their company, or what you promised them last Tuesday on the call. It wrote a lovely email for a person who basically doesn’t exist.

That right there is the moment most people quietly decide AI is super overhyped and useless as long as it’s not deeply integrated into their business, and I cannot blame them honestly.

But think about it for a second: you asked the most capable tool you’ve ever touched for help with your actual work, and it handed you a fortune cookie.

Here is what is really going, and it took me, for some reason, embarrassingly long to see it.

That AI LLM model isn’t simply dumb, as you might think. It’s just blind. Imagine that it has read and elaborated what feels like the entire internet, and it knows literally nothing about you. Not your inbox. Not your clients. Not the doc you’ve had open for 3 days. Every new chat starts from absolute zero, like a brilliant new hire who somehow forgets the entire company overnight and shows up needing the full tour again every single morning.

You’re just reexplaining over and over again. That’s not how a system like this should be used. The potential is huge but people are using it only at 5%.

Subscribe now

So you adapt without even noticing you’re doing it. Before you can ask for anything good, you paste the client call, the client history, your brand voice, some numbers from the last call. You type an entire page of explanation to your AI agent today and tomorrow you do the whole briefing again, from scratch, because there’s no memory.

That isn’t using AI. That’s being its assistant who hands the right context over and over again.

I lived like this for months myself. I’d ask for content ideas for my X account last year and I’d get “5 tips to boost your productivity,” because how would it possibly know I write about calling out AI hype? It was guessing at my life from the outside and silently making things up.

I thought this could not go on anymore and I thought: what if these models would have the right context every single time I ask them?

Then I found this awesome thing called connectors, and the guessing basically stopped.

Here’s a super simple explanation.

A connector is just a small switch that lets the AI look inside one of your real apps.

You want your AI to craft high quality posts based on some docs you have, you can connect Notion to ChatGPT or Claude and it automatically sees the necessary pages and folders.

You want to know what email updates you have for the day or you need to send a quick email to your colleague asking for a remote meeting, you connect Gmail.

Maybe you use google docs or spreadsheets for handling your client call summaries, you connect your Google Drive and the proposal creates automatically, you review the pdf and send it through the Gmail connector directly to the client. Easy peasy, right?

Basically, instead of you describing your work to it, it just reads the actual thing. And when you allow it, it writes back. The real note, the real reply, the real row in the real file. No copying, no pasting, no morning briefing.

The first time I connected my Notion and typed “what’s still sitting in my Substack content database that needs review” and it just listed them. I laughed out loud because first of all it felt like magic and because it was so obviously how this should have worked from day one.

Now the part nobody selling you this will say out loud.

I got excited and connected a lot of tools. 14 apps in a single Saturday. Notion, Gmail, Google Cal, some file thing I haven’t opened since, Granola, etc. By the next weekend I had deleted 9 of them.

Because more connections is not more power. The internet insists otherwise of course. Every other post screams “the 10 connectors you NEED” or the “15 ultimate automation tools that will boost your biz 50x”. That makes me laugh and reflect at the same time because people are actually integrating more and more tools every day thinking that is the solution for them being more productive.

I do deeply believe that less is more. Your day should be all about spending more time on the work you actually care and really want to do, and stop wasting time on doing the boring, manual stuff.

So for the connectors, you’re supposed to use only the ones that actually make your day-to-day less stressful, easier to manage and with which you can get your work done faster.

For example: why do you need to spend 3 hours crafting a proposal while you can do it in less than 10 minutes by using a google drive connector and the client calls you did for the past 2 weeks. Imagine, you could spend that remaining time to outreach to other clients instead, focus on sales or service delivery.

Coming back to why less is more, I have 2 main reasons to connect less stuff.

First, every connected app quietly hands the AI a menu of everything it can do, before you’ve said a word. A couple of those is helpful. Fourteen is a pile it has to dig through to get to your actual question.

Second, every connection is a door into your stuff, and doors swing both ways. If the AI reads a sketchy webpage or a file with sneaky instructions buried in it, a connector you forgot you even turned on can get used in a way you never meant. The people getting real value out of this are not the ones with the most connections. They’re the ones with the fewest and that are working on purpose.

So I kept 5 of them: Notion, Google Drive, Google Calendar, Fathom and Apify. Because that’s the room I actually live in all day. My calendar, my client notes, my half-finished article drafts I come up with when I’m feeling creative.

Did connecting it make me 2-3x times faster? I’d say more yes than no. The thing is now I’m focused on the actual core tasks of my work, not preparing onboarding and proposal docs for 3 hours straight. Anyone who tells you it skyrockets your productivity is probably selling you a course. The honest win is smaller and better than that. I basically stopped being the guy that does it all. The thing that used to be twenty open tabs and a lot of frantic clicking is now one sentence in a chat with a running Claude skill.

If you want to try it, don’t overthink it. Open your AI app’s settings, find the section called Connectors, and pick the single app where your real work lives. Click connect, log in when it asks, approve. That is the entire setup.

Look, I’ll make it even easier for you, I’ll guide you through it with the example of the Claude app:

1. Open up the Claude desktop or web app and find the “Customize” element on the left

2. Click on “Connectors” here:

3. You click on the plus button. Yes, I’m testing other connectors right now (n8n for automations and Lucid for charts):

4. Finally search for the connectors you need

3 easy wins you can steal today

Connecting one app is nice. Here’s what to actually do with it. Three things I run almost every day, the exact prompt, and roughly what each one saves me.

Win 1: Draft your content without briefing anyone (Notion)

Connect Notion with the method I showed above, then open a new session in Claude Cowork (cowork > normal claude) and type something like:

Create three assets in my Notion workspace:

## 1. Content Framework
Build a reference document that defines:
- My ICP (Ideal Customer Profile)
- My target audience
- Their pains, goals, objections, and desires
- Core content pillars
- Brand positioning
- Key messaging
- Attention psychology principles
- Hooks that fit my audience
- CTA framework
- Topics to avoid
- Content constraints and rules

This document should become the source of truth for every piece of content.

## 2. Voice Guide
Analyze my existing content, including my Substack Notes, and create a writing guide that captures:
- My writing style
- Sentence structure
- Tone
- Vocabulary
- Common phrases and filler words
- Formatting patterns
- Hook style
- Storytelling style
- CTA style
- Emoji usage (if any)
- Things I consistently do
- Things I never do

Include examples from my previous posts and Substack Notes so future content sounds like me rather than generic AI.

## 3. Content Database
Create a content database with the following properties:
- Title
- Platform
- Status (Draft, To Review, Scheduled, Published)
- Published Date
- Published URL
- Content Type
- Topic
- Content Pillar
- Draft
- Notes
- Source Material
- Last Updated

After creating the Content Framework and Voice Guide, populate the Content Database with an initial week of content (7 Substack Notes). Each note should:
- Be aligned with my ICP and content pillars
- Match my writing voice
- Explore a different idea or angle
- Be saved with the status "Draft"
- Be ready for review and publishing

It pulls from my real notes and style, so the draft sounds like me, not AI. It saves me about 3 hours every week by eliminating the need to re-explain what I write about.

Win 2: Turn a call into a follow-up before you forget (Fathom)

Connect Fathom (free video call + notes transcriber) to Claude.

Right after a client call:

Pull my last call with [client] from Fathom, list the action items, and draft a follow-up I can drop in an email.

The summary and the message write themselves while it’s all still fresh. Saves me around 30 minutes per call, and I stopped letting follow-ups slip entirely.

Win 3: Clear your work inbox in one pass (Slack)

First thing in the morning:

Show me the messages from the last 24 hours that actually need a reply, and draft a response for each that I can edit.

You go from staring at 40 unread to editing 6 drafts. Saves me 30 to 45 minutes a day, plus a chunk of low-grade dread.

Notice the pattern. You’re not learning new software. You’re just telling the AI to go look at the thing it can finally see.

Two things actually worth remembering. The AI only ever sees what your own account can already see, so if it can’t find something, you’ve connected the wrong account, not hit a broken app. And don’t let it act unsupervised. When it offers to send or delete something, it should ask you first. Read that little prompt before you click yes. Looking at your stuff is harmless. Sending an email as you is a different thing entirely.

That’s it. One app, wired to the place your work lives. Sit with that for a week before you even think about a second.

You don’t need a wall of glowing connections you’ll install once and never use again. You need to stop briefing a stranger every single morning.

Connect the one that ends that kind of monotonous work and simply ignore the rest.

One quick last one, because it actually pays for itself.

I used to spend about 3 hours on a single client proposal. Digging through call notes, copying numbers, formatting, second-guessing every line.

Now it takes me under 10 minutes. The Google Drive connector reads the call notes from the last two weeks, drafts the proposal in my own template, I review it, and it goes out the door.

That’s not a productivity flex. It’s 2 hours and 50 minutes back per proposal, time I now spend on outreach and actually delivering the work. Across a busy week that’s the whole difference between chasing clients and serving them.

The result is simple. The setup behind it is the part worth paying for. Below the line is the exact Claude skill and the prompts I use to turn two weeks of messy call notes into a finished, sent proposal in under 10 minutes. Copy it, plug in your own template, run it today.

The 10-minute proposal skill

Here’s the whole thing. By the end you’ll have a repeatable setup where you say one sentence and a client-ready proposal comes out, built from your real calls, in your real format.

I can also share a template I use.

Step 1: Put your raw material in one place.

Drop your call notes, transcripts, or Fathom summaries into a single Google Drive folder for each client. The AI is only as good as what it can reach, so give it the real notes, not your memory of them.

Step 2: Save a proposal it can copy.

Take your best past proposal, the one that actually closed, and keep it in that same Drive as your template. This is what makes the output sound like you instead of generic. No template, generic proposal. Good template, YOUR proposal.

Step 3: Run the prompt.

Paste this, swapping the brackets for your own:

You are my proposal writer.
1. Read my call notes with [CLIENT] from the last 2 weeks in [DRIVE FOLDER].
2. Read [MY TEMPLATE PROPOSAL] and match its structure, tone, and section order exactly.
3. Draft a proposal for [CLIENT] using their actual words, goals, and numbers from the calls.
4. List any assumptions you had to make at the very top so I can fix them.
Do not invent pricing or scope I never mentioned. If something important is missing, ask me instead of guessing.

Those last two lines, flag your assumptions and don’t invent, are the difference between a draft you trust and one you have to fact-check word by word.

Step 4: Review, then send.

Read the assumptions list first, fix those, skim the rest. Then: “Looks good. Export it as a PDF and draft a short, warm email to [client] with it attached, referencing our last call.” The Gmail connector writes the email, you hit send. Keep send on ask-first, like we said earlier.

Step 5: Make it one sentence next time.

Once it works, save that first prompt as a reusable Claude skill, or just a saved prompt, called “Proposal from calls.” Now the whole job is one line: “run my proposal skill for [client].” That’s the sentence I mentioned at the very top of this article.

What to check: a good run quotes the client’s own language back to them and follows your template’s sections. If it reads generic, your template wasn’t specific enough or the notes were too thin. Fix the input, not the prompt.

That’s the system. Call notes in, sent proposal out, about ten minutes, and it sounds like you wrote it on a good day.

Before I let you go, I wanted to say thank you for arriving this far and I wanted to gift you quick little guide on how to master Claude Skills. Click on this link for more: Resource - Use Claude Skills like a Pro

Subscribe now

Talk soon,

Ilya

For 2 months I published every single week.

No junk posts, no ai slop, none of that. Posts on topics I actually wanted to write about, that I’m competent enough to talk about. Same day every week, like clockwork. By every rule you get handed, I was doing it right.

And my subscriber count just sat there.

You know that feeling? You’re not lazy, you’re showing up every single week, and the graph stays flat while you quietly wonder if you’re just not good enough.

I was being consistent. And here’s the part nobody tells you when they hand you the “just be consistent” line: consistency only moves the needle if you’re also iterating on the process behind it. Showing up isn’t the same as improving.

So I did what everyone does. I blamed my writing. I tightened hooks. I studied structure. I rewrote openings until they were sharp. The line stayed flat.

It took me longer than I’d like to admit to see the real problem. It wasn’t my writing.

Good content works on 3 pillars

Here’s the simplest way to think about it.

Anything worth publishing sits on three big pillars:

• What you like

• What you actually know

• What is trendy

Look at your last 5 posts. You’ll almost certainly find pillars one and two all over them. The stuff you enjoy, filtered through what you know, through the lens of your expertise. That’s most creators. That’s exactly where I was a couple of weeks ago too.

The third pillar is the one everybody avoids.

Think about it. The first two are free. They live in your head. Instant, always there, no extra work, no research. The third one is trickier. It lives outside you, in your readers, their comments, THEIR heads, the posts they’re actually engaging with across your niche. The only way to get it is to go out and look.

And going out to look is the part everyone skips. I know, because for 2 months I did it by hand.

Every Sunday, the same five competitor newsletters. Open the page. Scroll the recent posts. Read. Check what fit my framework. Paste the good bits into Claude for inspiration. A whole ritual, start to finish, by hand. Then one Sunday I asked myself the question I’d been avoiding. Would a real automation expert keep doing this by hand? I knew the answer. I was just tired. And tired is a terrible place to start writing from.

Because by hand, that’s all research ever becomes. You skim subject lines. You glance at what got engagement. You paste something into a doc and tell yourself that’s research. It’s pattern-matching that resets to zero every week and never gives you the same answer twice.

So here’s what quietly happens. When research is painful, you skip it. Not on purpose. You just keep writing pillars one and two, straight from your head, into a feed that moved on without telling you.

It feels like authenticity. It just doesn’t grow your list.

There was an article this year where the author looked at 3,230 notes from nine creators in the AI and tech space. Three posted almost identical volumes, 475 to 499 notes each. Their per-note engagement averaged 8.7, 33.5, and 37.7 likes. Same volume. Same niche. A four-to-one gap.

The difference wasn’t how often they posted. It was whether their content kept tracking what the audience wanted, or quietly repeated what the writer already liked.

The creator at the bottom? Her best posts had nothing to do with the niche she thought she was in. Her audience had moved and her feed hadn’t noticed. She was standing on pillars one and two, writing from her head, while the third pillar sat there ignored.

That’s the real reason most Substacks grow slow. Not the writing. The missing third pillar. And the only way to put weight on it is to research your niche properly: what your competitors’ readers are actually responding to, which topics keep climbing, and which ones you keep walking past. Done right, that stops being a Sunday headache and starts telling you what to write before you write it.

That’s what the rest of this is about. Most people’s version of research is broken at the root, so let me walk you through the three stages of it, and the one that actually compounds.

Why most “automation” is still Stage 1

From what I’ve seen, there are 3 stages of niche research. Most people are stuck at Stage 1 and don’t know there’s anything past it.

Stage 1 is basic scraping. You, or a tool, pull recent posts from a few competitors. You read them. You hunt for patterns. Better than nothing. But it resets every single time. Each session starts cold. No memory of last week, no comparison against what you’ve already covered, no accumulation. You gather more and more, it turns into noise, and you walk away with less clarity than you started with.

Stage 2 is context-aware research. An agent that knows who you are before it starts looking. Your content pillars. Your audience’s actual pain points. The angles you’ve already written. The topics you’d never touch. With that sorted, the output stops being “here’s what’s trendy” and becomes “here’s what’s worth writing, given who you are.” Same raw input. Completely different answer.

Stage 3 is research that compounds. The agent doesn’t just scrape now. It remembers what it found last week. It compares this week against the last four. It notices you ignored the same gap twice. It tells you what changed since the last run. The fourth scan is sharper than the first because every run adds to the picture.

Almost everyone I know is at Stage 1, or they don’t do research at all. A few stumbled into Stage 2 without naming it. Stage 3 is where the real edge is, and it’s where I ended up.

I built a Stage 3 version using Claude Cowork and one MCP connector. Yeah I know, sounds complicated, but it really isn’t. Bear with me and I’ll fully walk you through. Here’s what Stage 3 gives you concretely.

What Stage 3 actually feels like

The week after I set it up, I sat down on a Tuesday morning, typed one sentence into Claude, and went to make coffee. 10 minutes later there was a research brief sitting in my project folder telling me:

• which topics were heating up across the newsletters I track

• which posts actually overlapped with my audience

• which titles and hooks pulled the most in the last 30 days

And at the bottom, one quiet line: “You’ve skipped gap X in your last 3 articles.”

That last line is the whole thing. That’s the agent learning from its own history instead of starting cold.

Not “here are ten content ideas.” Ten ideas is just more noise.

It had enough memory to notice I kept walking past the same opening, week after week, and it put that in front of me. That’s closer to having an editor than a research tool.

The Sunday before, I’d burned two hours and still felt unsure what to write. The week after, I had more useful signal in 10 minutes than a month of manual sessions had given me.

The agent didn’t improve my writing, but I just stopped guessing about that third pillar.

How it actually finds the gaps

When I first describe this, people assume the “find me an opportunity” part is magic, or worse, fluff. It isn’t. It’s boring set math, which is exactly why it works.

Stay with me here, this part matters.

Every run, the agent has four things loaded about you: your content pillars, the topics you’ve already covered, the topics you refuse to write about, and what you’ve defined as a gap worth your time. Then it pulls every competitor post from the last 30 days with the title, full text, and engagement numbers attached.

A real opportunity has to clear four filters at once:

• It sits next to one of your pillars, not off in some random niche

• It’s pulling actual engagement right now

• You haven’t already written it

• It’s not on your “never touch” list

The topic that survives all four is the gap. That intersection is the entire trick. No vibes, no “the AI just knows.” It’s the third pillar, made visible.

Then the research log adds the one thing a single scan can’t see: time. Because each run writes down what it found, the next run can say “engagement on this angle has climbed three weeks straight” or “this is the third scan in a row you’ve ignored this.” A one-off tool can’t do that. A system with memory can.

That’s the difference between a scraper and an editor.

What this is actually made of

Here’s the whole architecture. Four pieces. Knowing them upfront is enough to start building your own version, but it’s not required.

1. The scraper (Apify). Pulls the recent posts from each newsletter on your watchlist. Titles, full text, engagement metrics, publish dates. You set the list once and it runs on demand. The actor we’ll be using is called fatihtahta/substack-scraper, and it pulls straight from Substack’s public data. For five newsletters it usually costs around $0.60 a run and finishes in under 90 seconds. The exact cost moves with how many posts you pull and how much full text each one carries, since you pay per result. A heavy week might be a dollar, a light one thirty cents. Apify’s free plan gives you $5 of usage every month, which covers this comfortably.

2. The context layer (four plain-text files). This is the piece most people skip, and it’s exactly why their AI research comes out generic. Four highly specific markdown files that tell the agent who your audience is, what you’ve already covered, what you avoid, and what counts as a real gap for you. Skip this layer and every brief comes back as the same trending topics your competitors already wrote.

3. The orchestration layer (Claude + a skill file). One instruction file that tells Claude what to do, in what order, and what the output should look like. Claude reads your context, calls the scraper, and runs everything through your editorial lens.

4. The output and memory layer (two files in your project folder). Each scan drops a dated brief file into your folder, and appends one line to a running research log that lives right next to it. No second app, no database to name correctly. That log is what turns this from a one-time tool into a system that compounds.

Those four pieces are the whole thing. The section below shows exactly how to wire them together. No code. About 30 to 45 minutes of your time to set it up once. After that it runs from a single sentence, or on a schedule.

What you get below this line

The first half gave you the diagnosis: slow growth is almost never a writing problem, it’s the missing third pillar.

Below, I walk you through building the system that fixes it for your own niche. No code. No engineering background. Every file, every prompt, every step, including the messy parts. Like what to do when the first scan comes back looking generic, which it will, and how to fix it in 5 minutes.

This is the exact system I run every 2 weeks. It costs about $0.60 a scan. I’m not going back to doing it by hand.

Let’s build it.

I know it looks technical at first glance. It isn’t. I’ll walk you through each step like you’ve never done this before, because the first time someone told me to “just configure your MCP connector,” I had no clue what that meant either.

So bear with me, ok? Here’s what we’re building: a research agent that lives inside one Claude Cowork project, knows your newsletter cold, scrapes your competitors on command, remembers what it finds over time, and drops a clean brief straight into your project folder whenever you ask.

You need three things:

• A Claude account (the $20/month Pro plan is fine for this)

• An Apify account (free to start, Apify gives you $5 of usage a month, costs cents per run)

• About 30 to 45 minutes, most of it spent on your own context

That’s it. No database setup, no second app to keep in sync. Now let’s dive in.

Step 0: Set up your project (do everything here)

This is the step that makes or breaks the build, and the original version of this guide skipped it. Do not run these steps across a bunch of separate chats. The whole thing lives in one place.

In Claude Cowork, create a new project called something like Substack Research. Connect it to a real folder on your machine.

Read more

I got my tenth rejection on LinkedIn the other day and went to make coffee.

No stress. No existential crisis. Just coffee.

I remember thinking: this is the part they don’t put in their highlight reels.

The part when everything seems to work but nothing is clicking. And yet, you’re still trying. Showing up and trying not to suck.

The version people see

I saw a specific image of entrepreneurship sold online. The freedom. The laptop in a café. The “I work for myself” line at parties. The so-called “passive income” that isn’t actually passive and you’re just drowning in work 12 hours a day.

It’s not a lie, exactly. But it’s a highlight reel of maybe 5% of the entire experience.

The other 95% is rejection nobody sees. Conversations that go nowhere. Friends slowly drifting because they suddenly find you “different.”

When someone tells me “You’re different,” I say: “Good. It means I failed, sacrificed, learned. I’m not the same person you last saw.”

A lot of people switch up on you when you start changing. It leaves a specific kind of bitterness - hard to describe to someone who hasn’t felt it. You’re growing, something’s moving, but you’re not there yet. Your current friends don’t hold your attention the same way, and you don’t have new ones who’ve already done what you’re trying to do.

I also noticed that when you start building, the commentary starts almost immediately.

The friend who “just wants to be realistic.” The family member who’d much rather see you in a stable job than roaming in the unknown for years. The classic Joe at the bar who says your startup idea already exists, nobody needs it, it won’t work - not because he’s thought about it, but because he doesn’t understand it, or because he wants to keep everyone else at his mediocre level.

I learned to do a couple of things early on and they truly changed how other people perceived me - and my inner peace as well.

The first one is to stop talking about it.

Not out of shame. Just out of life strategy. Talk less about your goals, your struggles, your wins - and people will genuinely think you’re like them. Meanwhile you’re quietly building something. And one day, when it’s undeniable, they’ll call you lucky.

Success becomes the only real proof.

The second one is to stop listening.

Everyone talks so much. 99% of people don’t have enough context about your experiences, your fears, your goals - so why would their opinions help you? They won’t. The only person with enough context about your life is you. Not Joe. Not your best friend. Not your girlfriend. Not even your mom. You.

The loneliness is real

Here’s what happens when you stop talking and stop listening: you enter this bubble where the real fight happens within you. You fight for clarity, fight against your poor habits, the detrimental thoughts that always sabotage you right when things start moving.

You start on the best self-improvement journey of your life, trying to become the best person in your lineage. The best in your surroundings. But it comes with a price.

You also start realizing there aren’t many people left you can actually talk to. Not many who get you, your goals, the pain you’re going through.

Those aren’t bad people. They just can’t meet you where you are.

Mediocrity isn’t the answer for you either. So you stick to your plans. And somewhere along the way you realize you’re alone on this path. Like, literally alone.

You’re lucky if there are one or two people you can share things with. But even they won’t fully understand what you’re going through.

What breaks people

That loneliness follows you into the professional side of things too. When I started, I didn’t think it would be easy. But I didn’t expect it to be this hard.

I don’t know how many times I’ve been told no since I started. Ignored, dismissed, ghosted. Told my thing was useless by people who hadn’t even looked at it.

It sucks. It demoralizes you and makes you rethink this life decision: is entrepreneurship really for me?

I push that thought away and move on. Because sitting there overthinking every life decision is just being stagnant out of fear of failing. What I do is simple: try, fail, iterate. There’s no secret formula for improvement. Just one word. Mistake. We change our behavior based on the mistakes we make and what they produce.

But the famous question always comes up after: what if it doesn’t work out eventually? Well, I’d answer with two quotes I heard once:

First - if you truly try your best, you won’t fail because at least you’ll learn.

And second - you’re overestimating what you can achieve in one year and underestimating what you can achieve in a decade.

But here’s the thing - it’s not rejection that actually breaks people. People can absorb rejection. You adjust, move on, try again.

What filters people out is the ambiguity. You don’t know where to start. You listen to people who seem to know more than you, adopt their strategy, then realize it doesn’t fit your situation - worse, that it was never aligned with your inner values to begin with.

You hesitate. You try. You fail. Nothing seems to work. Everything is dark.

And then one person somehow becomes a client. One person says yes. And that one moment gives you just enough to keep going.

The pressure isn’t any single one of these things. It’s carrying all of them at the same time.

It’s not for everyone

It for sure isn’t for everyone. The internet sells “anyone can do it” because it gets views and sells courses.

My honest answer is: this path requires a specific kind of person. Not a superhuman. Just someone wired differently - for rejection, for ambiguity, for failing publicly and showing up the next morning like it was just data. Someone who doesn’t wait for the fog to clear before moving. Someone who can sit with “I don’t know if this will work” for months, maybe years, and not let it eat them alive.

They make it work anyway.

Most people aren’t built that way. And they shouldn’t force themselves to be.

Talk soon,

Ilya

I used to type the same 200-word instruction to Claude three times a day.

Same tone. Same format. Same context. Every. Single. Time.

It felt productive. Really, it was just manual labor with a fancier interface.

Most people using Claude have never heard of SKILLs. They’ve been at it for months, still rebuilding from scratch every single session. There are two completely different ways to use Claude. Almost everyone is stuck in the first one.

Chat mode vs. Cowork mode

The first way is what everyone starts with. You open a new conversation, explain what you want, get something back, refine it, repeat.

That’s chat mode. It works. But every session starts from zero. Claude has 0 context, zero information to start working with.

Claude knows nothing about you, your goals, your rules, or how you like things done - until you tell it again. And again. And again.

Feels repetitive, right? Yeah it does.

Well, there’s a second way, which is completely different.

Instead of explaining yourself every time, you configure Claude once - your context, your voice, your rules - and then it just knows. You trigger it, it works. No briefing. No back and forth.

That’s a SKILL.

A SKILL is a saved configuration file Claude loads before it starts working. Think of it as the difference between hiring a new freelancer every morning and actually training someone who works with you long-term.

The SKILLs I actually use

These are the SKILLs I actually use in my day-to-day or weekly work. I post on Substack, but everything here applies if you’re on LinkedIn, X, or any platform where you show up consistently.

Content planning. Every week, instead of staring at a blank page wondering what to write, I trigger one SKILL and get 14 post ideas - grounded in my voice, my top performers, and my content pillars. No briefing. No explaining who my audience is. It already knows.

Article drafting. I write the rough idea, the angle, the key moments - all of it comes from me. The SKILL takes that and builds a structured expansion in my format. Think of it as a thinking partner that knows your voice, not a ghostwriter. The angle, the story, the opinions - all mine. The SKILL handles the scaffolding so I’m not rebuilding structure from scratch every time. Phase 2 of this article for example went exactly that way.

Article review. Before I publish anything, I run it through a review SKILL. It doesn’t give generic feedback. It checks my draft against my own framework - my content pillars, my style rules, what to cut. It also looks for grammar and stylistic errors. It’s like having an editor by your side who has read everything I’ve ever written.

Lead qualification. I paste a list of prospects, usually a CSV, but an Excel file can work as well. The SKILL scores them against my ICP without me explaining what my ICP is. It already knows. I get a qualified list in seconds instead of going back and forth clarifying what I’m looking for.

Humanizing drafts. One trigger strips AI writing patterns from any text. No instructions needed. No “make this sound more human but not too casual but also not too formal.” It just does it. It literally removes the famous “It’s [X], not [Y]” pattern that you probably spotted in some AI-written content out there.

Five SKILLs. Each one built once. Each one running every week, either when I need it or on schedule.

Most content about SKILLs and automation targets businesses. Agencies, ops teams, enterprise workflows. But the people who get the most out of this are solo creators posting consistently. If you’re on Substack, that’s you. You’re already wearing every hat - writer, editor, strategist, distributor. A SKILL doesn’t replace any of those roles. It handles the clerical version of each one, so you can spend your time on the part that actually requires you.

What’s actually inside a SKILL

Well, whatever you actually put inside of it. Every SKILL has three things.

Your context - who you are, what this is for, who the audience is.

Your rules - how Claude should behave, what format to use, what to avoid.

Your trigger - what kicks it off and what it expects from you each time.

The key insight here: you’re not writing a better prompt. You’re documenting how you think about a task, once, so you never have to explain it again.

The first time you set up a SKILL, it takes some thought. You’re essentially writing a job description for someone who never forgets anything and never has a bad day. That’s the work - straight up.

Here’s what a simple one actually looks like. Say you want a SKILL for drafting Substack notes:

name: draft-note
description: Draft a Substack note in my voice from a rough idea.

Rules:
- Short paragraphs, one idea per line
- No em dashes, no emojis, no hype language
- Open with a real hook, not a lesson
- End with a grounded takeaway, not a generic CTA

That’s the whole file. Claude reads it before it starts, follows the rules, and uses whatever context you give it that session. One file. Infinite runs.

Triggering it is even simpler. You open Claude, type /draft-note, add a sentence or two about your idea, and press enter. Claude reads the SKILL, follows the steps, and produces output that already fits your format. No briefing. No re-explaining your tone or your audience. The first time it works, it feels almost too easy.

What makes a good SKILL

Keep each SKILL tight. One task. One context. The moment you try to make it do everything, it does nothing well.

And here’s the thing that stops most people before they even start - they assume they need to code or write in some special syntax. You don’t. If you know your process well, write it yourself in plain language. If you’re not sure how to structure it, just describe the workflow to Claude and ask it to build the SKILL file for you. Either way, you end up with a file you own, can edit anytime, and reuse forever. No technical knowledge required.

Where to start

If you write content of any kind, start with a content analysis SKILL.

You already know your niche. You already know your audience. You probably have a sense of what topics land for you. But reviewing your content performance every single week is a real burden. At least it was for me.

A simple rule: if you’ve explained the same process to Claude more than twice, it should be a SKILL. The upfront time is 15-20 minutes per SKILL. After that, it runs in seconds. Build one, use it a few times, and you’ll immediately know what to build next.

Here’s what to do:

1. Audit your process - how you review content and what criteria you actually use.

2. Write down all the instructions in plain language once

3. Craft a high-quality content framework in Notion or Google Docs. Your style, your pillars, your audience, your goals - everything that matters.

4. Let Claude reference that framework each session. Hand it your posts from the week.

Instead of typing: “analyze these last week posts”.

You get a SKILL that already knows your voice, your top performers, your content pillars - and produces ideas you’d actually write. One sentence input. Output that actually fits.

That’s your first hour of setup. After that, it runs on its own.

There’s a fully automated version of this that doesn’t even need you to copy-paste the posts. That’s a story for another time.

Where they sit in the Claude App?

This skills sit on your computer in a specialized folder but you can easily access them by following these steps:

You then click the plus button and either browse the available skills or create your own.

How to trigger one

Super simple. You just type /name-of-your-skill.

For example:

You then press enter or click on the arrow button and Claude starts working for you.

The honest part

Most people will read this and think “I should set that up.” Then they’ll go back to chat mode because it feels faster in the moment.

It isn’t. It just feels that way because the cost is invisible - spread across a hundred small sessions instead of one upfront hour.

The ones who do the hour upfront are the ones compounding. Everyone else is still copy-pasting.

Talk soon,

Ilya

I’ve been sold a lie. And look, if you’ve spent any time online trying to build a business, you’ve probably been sold the same one.

Building a real business is hard. Not the inspirational poster kind of hard where you’re forced to “embrace the grind” and everything works out eventually. Actually hard. You send 200 emails and get 3 replies. You spend three months on an offer and nobody buys. You show up every day and still feel like you’re standing still.

But nobody posts that. They post: “I’m making $10k a month working 3 hours a week.”

Yeah. Whatever.

I spent about a year doing exactly what I was told. Five sales courses. One coaching program. Every guru with a clean Stripe dashboard and a thumbnail that made it look effortless. I even sat down one evening and reverse-engineered a stranger’s payment numbers just to figure out what they were doing differently.

Looking back it was a pretty strange thing to do.

I kept thinking there was some sort of hidden truth I was missing. A system. The right email script. The right niche. The right funnel. The thing that would finally make it all click.

Nothing changed. Not because the tactics were wrong. Because I was pointed in the wrong direction.

Every few weeks I’d try something new. DMs. Cold emails. LinkedIn content. I’d give it a few weeks, see nothing happen, and convince myself I just needed a different approach. So I’d go find another course, another framework, another expert telling me they had the answer.

This went on for almost a year.

The Numbers Nobody Posts

Here’s what the “$0 to $10k/month” transformation story usually skips.

The 14 months before the $10k month. There’s the $2k month. The lucky $5k month. The $800 month when everything was silent. The -$350 month. The 3,000 cold emails before anyone replied. The 5 offers that flopped. The two clients who churned in the same week.

That content doesn’t get clicks. So it doesn’t get posted very often.

What gets posted instead and gets ton of engagement is usually: “I built one AI workflow and now it generates $120k/year replacing my entire marketing team.” Sure. But behind it there’s a business built over years, a pile of failed experiments they forgot to mention, and clients lost along the way. You’re watching the final cut. Not the footage left on the floor.

And here’s where it gets uncomfortable.

Subscribe now

I knew this. I logically understood it. But I kept going back anyway. Because when you’re stuck and frustrated and nothing is working, the highlight reel feels like proof that someone figured out something you haven’t. And maybe, if you just find the right thing to copy, you will too.

That’s the trap. I started to figure out that:

You can copy the what, but you can’t copy the why.

Let me explain. You can try copying what another successful entrepreneur or professional is doing on the internet, you might get that initial result in, but you’re not learning practically anything. You’re not understanding why X is working and Y isn’t. You’re just copying what another person told you without questioning why it works.

Tactics change, while the strategy behind them usually doesn’t. So instead of copying what someone does, I started studying why it works - the actual mechanics behind what’s driving results.

The Moment Things Started to Shift

Several weeks passed. I cleaned up my Instagram and YouTube feeds - cut out most of the success story accounts, the highlight reels, the “I scaled to X” posts.

And things suddenly got quieter.

I started reflecting on life more, paying more attention to what I actually cared about. I embraced critical thinking instead of just passively consuming hypey content. And the more I sat with it, the more something obvious started to surface - something I’d been ignoring for all these months.

Everything works. Like almost all strategies can give you measurable results. Cold email works. DM works. Content works. Paid ad works. Cold calling works. Every single one of these strategies has produced results for someone.

So why wasn’t any of it working for me?

It wasn’t the strategy. It was the direction. I was pointing all that effort somewhere that wasn’t mine - and no amount of effort fixes a wrong direction.

I tried cold calling for a while. I genuinely sucked at it. At first I told myself it was just uncomfortable, that I needed to push through. But I stopped kidding myself. Cold calling just wasn’t me. The way I communicate, the way I build trust, the kind of relationships I want with clients - cold calling didn’t fit any of that. It was like wearing someone else’s shoes and wondering why you kept tripping.

I was pointing all that energy in a direction that wasn’t mine. That was the whole problem.

The Direction Problem

Here’s how I started thinking about it.

Most people approach building a business like: put in more time, add more effort, get more results. That math checks out. Until the direction is off.

What I noticed working is actually this:

If your direction is off - if you’re copying someone else’s approach, using their voice, running their playbook inside your business - you can multiply that by all the time and effort you want, but ultimately you’ll end up very far down the wrong road, very fast. The compounding works, but it works against you.

And when you find your actual direction, your vector, the one that fits how you think and communicate and work - then time and effort start to compound the right way. Things start building on each other. And that’s when it begins to actually move.

Your vector isn’t something you find by watching more content. It comes from actually trying things, noticing what fits and what doesn’t, and being honest with yourself about the difference. It’s built from your specific experience, your personality, the way you naturally talk to people, the problems you understand from the inside.

Nobody else has that exact combination. That’s kind of the whole point.

That combination is also what AI can’t replace. Everyone keeps panicking about being automated away. But the thing at the core of your vector - your specific experience, your instincts, how you naturally build trust with people - that’s not something you can prompt your way to.

I’m Not There Yet

Am I where I want to be? Yes and no. I’m grateful for where I am, but nowhere near satisfied. I’m still figuring things out, still iterating, still getting things wrong more often than I’d like.

But something is different now.

Subscribe now

I used to grind because I was uncomfortable. Because I felt behind. All the results came from a sort of pain I had because somewhere out there, someone was posting their $10k month and I wasn’t. That kind of fuel burns, sure. But it burns you with it.

Now I grind because I actually enjoy it. I enjoy the process. I enjoy diving into the work, even though it’s not satisfying at its fullest all the time and it doesn’t have to be. Because the work I’m doing is the work I chose - not the work some guy on YouTube told me to do, not the playbook someone sold me for $997, not the strategy everyone else is copying. Mine. Built around how I think, how I communicate, what I’m actually good at.

And that’s a completely different feeling. The same hours, the same effort, but it doesn’t feel like punishment anymore. It feels like building something that’s actually yours.

Chat soon,

Ilya

Everyone is talking about how great Claude is. And every time I ask someone what they actually use it for, I get something like this:

“I write emails with it.”

You send a message. It replies. You send another. It replies again, this time with a little encouragement at the end. You feel productive after the chat. You close the tab.

Congratulations. You just paid for a very empathetic search engine.

Here’s what nobody talks about: Chat mode feels productive. You asked something. You got an answer. Brain says: task done.

But that feeling is exactly what keeps most people stuck.

Claude is not a search engine you talk to. It’s closer to a contractor you can deploy on actual work. The difference between using it to draft an email and using it to run your back office is not a skill gap. It’s a mindset gap.

There are three ways to use Claude. Most people never leave the first one.

Chat - The base

This is what Claude welcomes you with: a chat window where you write a question and it outputs an answer.

Useful for drafting an email, summarizing a document, getting an answer at 2AM when you cannot remember the difference between affect and effect. Nothing wrong with this. I used it this way for weeks before realizing I was barely scratching the surface.

But you are still doing all the heavy thinking yourself. Every prompt needs your attention. Every output needs a review. You are still the judge deciding whether it’s good enough.

Claude here is a fast helper and nothing more. If this is all you use it for, you are essentially renting a Porsche 911 to do grocery shopping.

Cowork - The delegate

This is where things get genuinely interesting.

Cowork is the second option on the desktop app, so switch to it now.

You operate inside project. This is how you create a project and attach to a folder on your computer:

You connect it to your files and tools via Connectors, and it stops answering questions and starts doing the work for you. Client work in one project. Substack content in another. Each project carries its own context, its own connected services, its own history.

But the real unlock is Skills.

Skills are pre-built workflows that give Claude specific capabilities. A proposal drafted from a meeting transcript. A lead qualification table from a LinkedIn list. A weekly Substack engagement report ready by Sunday morning. Not prompts you rewrite from scratch every time and you chat back and forth on. Workflows you build once and trigger on command.

You describe what the workflow should do. Claude follows it every time. No variation, no re-explaining yourself.

Skills combined with scheduled tasks? I’ve automated roughly 30% of my content analysis work this way. The tasks are still there of course. I just stopped being the one doing them manually.

Then there is Dispatch. This is a nice tool if you don’t have access to your computer. So what it does is it pairs your Claude mobile app to your desktop. Send a task during your morning walk. Come back to something already finished on your screen.

At some point you stop being the person doing the work. You become the boss who says: “looks good, send it.”

Code - The builder

Let’s say you want to build something. A tool. An internal system. A product you have been sketching in your head for months.

Two years ago, that meant learning to code or hiring someone who could.

Now you use Claude Code.

You describe what you want to exist. A website, a mobile application, a script that automates a workflow. Claude scaffolds it, writes the logic, connects the pieces, and hands you something working.

This is how non-developers ship internal tools they could not afford to outsource. How small agencies build client-facing systems without a full engineering team. How solo founders automate workflows they would have previously just worked around.

Ok. Enough framing. Here is what this actually looks like in practice.

Use case 1 - The customer support system

We recently signed a project with an e-commerce company that needed a customer support automation for writing emails based on a set of predefined documents. Small support team, growing inbox, writing the same categories of replies manually every single day.

The deal started with a 45-minute discovery call. In the past, that call would mean two hours of notes cleanup and proposal writing afterward. This time, the recording went into Cowork. n8n pulled the video, read the transcript, identified the discussed scope, the pain points the client described in their own words, and drafted a full structured proposal. Into Notion in under ten minutes.

The client saw a proposal that matched exactly what they said they needed.

We signed that week.

Use case 2 - The weekly report you actually read

Every Sunday for weeks I was sitting down for three hours going through my Substack notes manually. What got likes. What got replies. What flopped completely. Opening each post, copying numbers into a doc, trying to spot a pattern.

3 hours. Every week.

Then I built a reporting Skill in Claude that does it for me. It pulls all the engagement data automatically, cross-references what I published, and shows what is working based on engagement, likes, and a set of predefined benchmarks. More platforms are getting added and Threads is next.

Now it takes three minutes. In Notion by Sunday morning. I just open it over coffee.

Use case 3 - The lead machine

A B2B team at a SaaS was spending two full days a week manually vetting inbound leads. Detailed ICP in Notion, spreadsheet of names. Every lead meant opening LinkedIn, reading a website, making a judgment call, updating a row. About 80 times per week.

Now Claude takes the ICP, checks each lead against their LinkedIn and website, and returns a scored qualification table. Qualified, borderline, not qualified - short explanation for each.

Two days became under an hour. The judgment is still there. The clicking through LinkedIn 80 times is not.

Still thinking about that team. Two days a week, every week, going through the same motions. Not because they were bad at their job. Because no one had handed them a better one.

Most people are not stuck because they cannot learn this. They are stuck because Chat feels like enough. It is fast of course, it is easy to use, and it gives you something to show for the last twenty minutes.

But there is a version where you stop filling the week with tasks and start finishing the quarter with results. That gap is smaller than it looks.

If you want to get there, subscribe below.

If you want to learn how to build your own Claude Skills, comment “skill” below and I’ll walk you through it.

Chat soon,

Ilya

You open LinkedIn on a Tuesday morning.

Three posts in, someone’s already made $40k using a “simple AI workflow” they built over a weekend.

By post six, a 24-year-old founder just replaced his entire ops team with Claude and saved $200k a year.

By post ten, you’re wondering what you’ve done with your life.

That’s the AI comparison trap. And it’s working exactly as designed.

The Contradiction Nobody Says Out Loud

The vast majority of people posting about their AI wins are the ones teaching it in their Skool community without ever actually implementing it in real business workflows.

They’re not ahead of you. They’re just better at looking like they are.

I’ve watched this pattern long enough to know how it works. You get a bold claim like “I replaced my $150k hire with a $20/month AI automation.”, zero to little context and the course link in their bio.

What they don’t mention: it took 3 days to build, no reliability, no results listed, and still needs someone to babysit it every week.

That’s not dishonest in a legal sense. But it’s not honest either.

The Guru Formula

I just noticed how AI content runs on a specific formula.

It usually goes as I said: bold claim, “clean” framework, course link in bio. What gets cut: the eighteen iterations that failed, the client who almost fired you, the weekend Claude almost deleted your entire database.

You see the finish line. Never the wrong turns.

And the finish line is designed to make you feel like you’re standing still.

What It Actually Looks Like

I build AI systems for real businesses. Lead qualification, content automation, customer support workflows, marketing pipelines.

Here’s what the process looks like: messy. Like, really messy.

Well, to begin with, the client doesn’t even know what they need. They just say something like:

I want AI and my sales pipelines automated.

You start with one suggestion. You listen. You empathize. Then you realize the pain is somewhere completely different from where you thought.

Everything you planned to build is suddenly wrong. You dig further, and it turns out the problem isn’t in their sales pipeline at all. It’s in their marketing.

An hour later you have a direction. You send the proposal. They sign off. You start building.

And here it starts looking even messier.

Version one doesn’t work. Version two kind of works. Version three works until something unexpected happens. Version four is something I’d actually show a client.

Nobody posts about this initial chaos or about versions one through three. Not because they’re hiding it. Because it doesn’t perform.

LinkedIn rewards the highlight reel. So that’s what gets made.

A Real Project

I recently worked with a B2B e-commerce company. They came in wanting one thing: a simple content workflow to save their team a few hours a week and boost their presence on LinkedIn, Facebook, and other platforms.

One discovery call. One and a half hours. That’s where we figured out what they actually needed versus what they asked for.

They didn’t need a content system. They needed a personalized database reactivation mechanism that would scale for their client base.

The proposal was sent. They signed off. Then the real work started.

Five to six days of planning before a single workflow was built. Not because it’s slow. Because if you skip that phase, you spend three weeks going back and forth on things you could’ve locked down in one session.

We mapped out three interconnected systems for their marketing campaign.

None of that was on the original brief. All of it became necessary once we understood the actual problem.

You won’t often see that version of the story on LinkedIn.

You’re Not Behind

The founders I’ve seen actually win with AI don’t spend much time watching what everyone else is doing.

They’re too busy on their own version one.

That’s it, really. Sit down with the actual problem, build through the mess, ship version four. Stop watching other people’s finish lines.

The gurus are optimizing for attention. You’re building something real. Those are not the same race.

Build Faster. Not Scroll Faster.

Next time you feel the pull to read another “AI saved my business” post, ask yourself one question: am I learning something specific, or am I watching someone else’s highlight reel?

If it’s the latter, close the tab.

You already have enough information to start. What you need is a version one.

Before you take any AI content at face value, just ask yourself three questions.

First: what did version one look like? If they’re not showing the failures, they’re most probably selling you the highlight reel.

Second: how long did it actually take? “Built over a weekend” is almost always fiction. The real timeline includes debugging, revisions, and something like two days where nothing worked.

Third: does this apply to your specific situation? A workflow built for a 10-person SaaS company is completely useless for an agency or an e-commerce business. Context isn’t a footnote. It’s the whole point.

If a piece of content can’t answer those three questions, move on. There’s real work to do.

PS: by the way, I’m running this contest now. The first 10 to comment “stack” under this post get the full No BS AI tool stack that I use in my day-to-day workflows to grow the agency. Worth bookmarking.

Chat soon,

Ilya

The Setup

So here’s the issue. Two weeks in a row I did the same thing.

Opened Substack. Copied last week’s notes into Claude. Said: “analyze these.” Got some generic output back. Closed the tab.

I kept telling myself the review mattered. And it did. But going through that process manually every Sunday was slow enough to make me frustrated every time.

The third Sunday I sat down to do it again, I stopped. Not to quit - to think.

What if I woke up the next Sunday morning with the review already done?

Unfortunately I had no way to pull the data automatically. No framework to assess against. No structured output to actually learn from.

So instead of fixing my habits, I fixed the system.

Subscribe now

The System

Simple idea. Three inputs, one output:

• Weekly Notes - every note I posted that week, with like count and timestamp

• Framework Doc - a Notion page defining my content pillars, hook rules, writing style, and what good engagement actually signals

• Weekly Report - one table: each note, the related stats (like count and comments), and what worked or didn’t

The framework doc isn’t 100% static. It updates when my thinking does - maybe once a month.

The Process

I spent a Sunday morning figuring out the pieces: what scraper to use, how to pull the framework doc, how to structure the final output.

Three stages ended up being enough.

The first uses Apify - a web scraping platform - to pull my recent notes automatically. Not just the text. The like count, the timestamp, everything I need without manually digging through the app.

The second uses Claude’s native Notion connector to retrieve the framework doc. No API key needed - just a one-time login. Claude reads the doc the same way it reads anything else you put in front of it.

The third is where it gets interesting. I built two lightweight custom skills inside Claude - one that handles the scraping, one that runs the review. Together they pass each note through the framework and ask the things I used to forget to ask:

• Did the hook open with a real moment or story?

• Is the angle specific or too broad?

• Does the voice hold throughout the note or does it drift into generic advice?

The whole thing runs with a single command. One skill calls the other, pulls the framework from Notion, runs the review, and writes the output back - all without me touching anything.

Here’s how the pieces fit together:

This is how it looks in for the week of April 13 → April 18:

What It Caught That I Missed

The first week I ran it, it flagged three notes I’d written off. Not viral - just quietly above my usual numbers.

All three had one thing in common: they opened with something that happened to me, not something I concluded.

No lesson. No angle. Just a short story.

This is an example of such a post:

I also noticed I’d been going more opinionated lately because it felt bolder and punchier. I felt like the quiet ones were too simplistic and not worth sharing. But what the data actually cared about wasn’t how confident the opener was, but whether someone stopped scrolling or not.

One of those notes had no take at all. Something went sideways, I kept going anyway, no moral attached. It hit roughly 3x my average that week.

The system also flagged hooks that were too broad. Not bad - just unfocused. I’d been trying to expand the post under different angles, multiple topics sometimes, but the framework kept saying the same thing:

Pick one, go deep.

It’s annoying to hear, but that’s usually how you know something starts to work.

Engagement → Before vs After

Before the system: 4-5 likes per note. Occasionally 10-11 if something really landed.

After two weeks: 7-10 minimum, 10-15 on average. Posts that connect are hitting 25-35, but I treat those as outliers.

I have almost no audience on Substack. These aren’t numbers softened by thousands of followers. Each like is a real person who chose to stop and read the note. When that doubles in a small room, it’s not the algorithm. It’s the content.

The Part Nobody Mentions

This only works because I already had a framework.

If I hadn’t written out my content pillars, my hook rules, my style notes - the reviewer would have nothing to compare against. It’d return a word count and a shrug.

The automation is maybe 20% of this. The document it reads is the other 80.

Subscribe now

Most people try to skip the document and go straight to the tool. They set up a scraper, ask Claude to review their posts, and wonder why the output feels hollow. And I’ll be honest, even with my current setup in 10% of cases it feels hollow.

It’s hollow because there’s nothing real to compare against.

Build the framework document first. Then automate around it.

The automation is nothing without something real to review against. Use AI to refine, not to generate the idea - that part still has to come from you.

If you want the full step-by-step to set this up, comment ‘notes’ below. First 10 people get it in their DMs.

Talk soon,

Ilya

I’m tired of seeing people using Claude like it’s Google.

Type a question. Get an answer. Close the tab.

Ok, that’s fine, but it’s like buying a sports car and never activating sport+ mode.

I build AI automations for a living. We built Claude-based systems into real business workflows, client pipelines, and internal systems. I’ve watched the same pattern play out with many clients: they start at Level 1, think they “get” AI, and wonder why it isn’t changing anything.

Subscribe now

The difference between Level 1 and Level 5 isn’t intelligence. It’s context.

Look, I’m gonna make it really simple for you.

Here’s how I use Claude at my agency - 5 levels of complexity, in order.

Level 1: The Search Replacement

This is the simplest and most straightforward way of using it.

You type questions. Claude answers them.

Useful? Yes. Transformative? No.

This is like googling on steroids and it’s usually where the basic 80% of Claude users stay. Not because they’re lazy. Because no one told them there’s anything above this floor.

Level 2: The Writing Partner

Now you decide to shift from asking to handing things off, like:

• “Draft this email.”

• “Rewrite this proposal in my voice.”

• “Write me a Substack post about X. Make it punchier, make this part softer…”

You give Claude your voice, your context, your goal. You become the editor and Claude creates you a structured draft.

This alone already saves you hours each week, but you’re still limited to your prompt.

Most people get here, think they know how to use AI and suddenly just stop improving.

But we aren’t satisfied by this, so let’s keep going.

Level 3: The Analyst

This is where the actual shift happens.

You inject more context.

You feed Claude real information instead of just raw prompts. Sales call transcripts. Support tickets. Six months of email data. That 40-page competitor report your team hasn’t had time to read.

Level 3 doesn’t just save you time. It identifies data patterns you wouldn’t be able to see by yourself. But here’s the thing I realized:

What unlocks level 3 even more is memory.

Claude remembers you across conversations - you attach PDFs, markdowns, documentation, images and more - pretty impressive huh. Here’s how it works in practice and how I use it:

• Connect Notion (or similar) to Claude via MCP.

• Start a new Claude project.

• Paste your files in → ICP doc, your clients csv, your tone of voice, etc.

• From that point on, every session starts with context already loaded.

You stop re-explaining yourself. Output already knows what you need.

We did this with a client building an AI customer service product. They had a raw lead list and an ICP document no one had time to touch. We loaded both into a Claude project, asked it to enrich and qualify each lead based on their LinkedIn bio and website, score by fit, and return a ranked list. It would’ve taken 5-6 hours to do that manually, while a Claude prompt with memory did that in under 15 minutes. No API. No code. Just context loaded once, reusable many times.

Level 4: The Operator

Now we get really serious. Claude doesn’t just think. It acts.

You connect it to the tools your business already runs on: Notion, your fave CRM, email, Slack, Google/Outlook calendars, Figma, etc.

We had an 800 lead list last month.

We ran it through Claude entirely through a well-crafted skill and asked to filter out leads that weren’t qualified under an ICP Notion doc we provided.

For each lead Claude scraped their LinkedIn profile, compared their bio to the ICP and voilà → 15 minutes, filtered lead spreadsheet in Google Sheets with a personalized starting message, ready to use for outreach.

Now as you know, this newsletter is called No BS AI, so I’ll be straight with you. What I described works. It’s not perfect every time and it requires fine-tuning - but it handles the kind of work that’s been sitting on someone’s to-do list for months, in under 15 minutes.

While level 3 gives you a smart advisor, level 4 gives you someone who does the work.

What unlocks level 4 is automation via SKILLs + scripts + scheduling (optional).

You adjust to run this automation each Sunday at 7 AM:

• Claude runs a /briefing SKILL every morning before you open your laptop.

• Pulls sales pipeline updates, flags urgent tasks in your CRM, summarises overnight activity.

• You sit down and it’s already done.

At this point Claude isn’t something you use. It’s something that runs.

Level 5: The Agent

Tools talking to tools. Outputs triggering other outputs.

Currently I’m working on this level myself.

At this level you’re not running one workflow. You’re running for example three at once.

Here’s a quick example.

I needed a weekly content briefing for Substack. Decided to set up 3 parallel research streams: my last week’s top posts, competitor activity, and trending topics in their niche. All three run simultaneously. Claude merges the results when they finish and delivers one briefing in Notion. What used to be two hours of manual research, drafting and review, now runs in 20 minutes.

This is not a just chatbot. This is a clean infrastructure.

Level 5 runs on parallelism. Multiple agents, each with one job, all working at the same time.

The key is of course mapping the full workflow before starting the build.

Time saved

This diagram shows you the rough amount of time saved for each level based on what I’ve seen at the agency.

The Real Gap

Look, I see so many people stuck at Levels 1 and 2, some of them reach Level 3 and very rarely I see people using Level 4 and above properly. Usually it’s not because they don’t understand AI.

They just never saw what’s possible and what’s not.

This newsletter is about cutting off the noise and delivering what actually matters to you.

AI gurus will tell you the key is better prompts, or writing more specifically, adding more context, more files to your knowledge base, more, more... Take a course and unlock some hidden layer of Claude’s potential. That advice has serious limitations.

Prompting lives at Level 1 and 2, where there is 0 leverage.

The real leverage kicks in when you stop prompting and start building reusable systems - context that loads itself, tasks that run without you. When Claude isn’t answering your questions but running your systems. When you don’t need to open the app at all.

Most people don’t know that version exists. They’re getting really good at Level 2, calling it AI fluency, while Levels 4 and 5 sit there completely untouched.

Here’s a summary of the 5 levels:

Most people won’t get past Level 3 this year. Not because it’s hard. Because they don’t know about Level 4 and 5.

Now you do. So start building.

Subscribe now

Talk soon,

Ilya

PS: if you enjoyed this article, I’d really appreciate if you shared it with someone who might be interested.

Share

Noise is your biggest enemy

The worst thing about building a business today is that everything actually works. Cold email works. Content works. Cold DM works. Every guru is right. That’s exactly the problem.

Last year I had a conversation with one of my followers on X, let’s call him John. I asked John what was stopping him from growing his business.

He wasn’t talking about skills, not even money. The answer was much more straightforward. He said:

There’s just too much information. I don’t know what to focus on.

He didn’t end up being my client or something, but I kept hearing that same answer at least 3 more times from different people since that conversation.

And they’d all say something like this:

“I don’t know where to start, I don’t know how to grow, I don’t know if this will work or not”.

And you know who was the first person I heard that issue from?

Me.

Yeah. Me. The person writing this. I was in the same spot. I was stuck reading stories, articles, how-tos, watching all those highlight reels, YouTube videos of somewhat successful entrepreneurs in the same niche as mine. They were successful and I wasn’t there yet. I wasn’t jealous. Just quietly aware of my potential and feeling like I was standing still.

In the end I had everything I needed on paper. The guides, the willpower, the obsession for the craft. What I didn’t have was clarity.

At that point in my life my #1 issue was: “How to get more clients?”.

Hear me out, I still have this issue, but at least it’s partially solved for now.

I had somewhat of a cold email kick for a period. Tried that. Built an automated CRM for dental clinics. Sent almost 900 messages. Had a terrible reply rate, like 4-5% and an even lower interest rate. Ended up not selling anything. I felt bad about the strategy, thought it didn’t work, thought my product wasn’t good enough. LIE.

Next, I tried content. Decided to post consistently on X for 8 months. It got me a couple clients, 1k followers. But I still thought it wasn’t good enough of a method. LIE. Again.

Finally, I switched to cold DM which stopped being popular when I started doing it. Sent thousands of DMs. Results? Well, the results were far better than the other methods that I tried, got me 5 booked calls in the first 2 weeks. That’s not the point though. Thanks to trying different things I realized:

There is no perfect strategy.

Everyone who’s telling you to use a specific method or tool or strategy, even they don’t know what’s best for you. And you know why?

Context.

The reason no strategy is universal is simple. These people don’t know your offer, your audience, your schedule, or how you handle rejection. You do. That context is yours alone, and no system built for someone else’s situation will fit yours perfectly.

What I actually figured out

Here’s what I saw happening. You follow someone online who swears by cold email. You set it up. Then another person says cold DM is the move now. You add that. Then someone tells you content is non-negotiable. You add that too.

Six weeks later you’re doing three things at 30% each, burning out, and wondering why nothing is working.

The tools aren’t the problem.

Cold email works. Cold DM works. Content works. Every single one of them has built real businesses.

You’re treating it like a buffet, when it’s meant to be a real commitment.

The gurus aren’t directly lying to you. They’re showing you their one thing and calling it a “universal truth”. They committed to cold email for 18 months and it worked. Of course they think it’s the answer.

But they left out the part where they ignored everything else.

Nobody says this part out loud unfortunately. So I will.

You’re adding new tools, new strategies, new shiny objects to your life only because committing to one thing means you might fail at one thing. If you spread across three different methods, you can always tell yourself it was the wrong mix. Staying scattered keeps you safe from failing at one specific thing. That’s it.

What I actually realized is that usually the noise isn’t coming from the internet. It’s in your own decisions. Every new tactic is just replacing the last before it had a chance to work.

Pick one channel. One approach. Give it 90 days of actual 100% effort, not half-effort while testing three other things. Then evaluate.

More information is not what you need.

You need to commit harder to less, not spread thinner across more.

So if you’re juggling three methods right now and wondering why nothing is clicking - that’s your answer. Everything works. That’s actually the problem. Pick the one that fits you best and stick to it long enough. You can adjust how you do it. Just don’t keep switching what you’re doing.

Talk soon,

Ilya

AI gurus never talk about this problem

2 weeks ago I was sitting down with a new client project.

Marketing automation pipeline for a company with both B2B and B2C customers coming in through their website and existing CRM. n8n, Supabase, a handful of API integrations. Nothing new on paper, but complex enough that one wrong connection would break the whole thing.

My old self would have opened n8n immediately and started laying out workflows.

It’s fast, sure. But not on complex builds.

I know because I did exactly that so many times. And every single time I paid for it. Not in money, but in hours spent refactoring the workflows or going back to change things I’d already built because I hadn’t thought through how they connected.

That specific frustration of realizing mid-build that the structure you chose was wrong, and now you have to unravel it.

I realized that the problem was never the code.

I lost 90% of the time on a build just because the architecture sucked.

I never sat down and planned the thing from start to end, at least on a high level.

So this time I stopped before taking any action. Just sat there for a moment at my desk and asked myself: what actually needs to happen before I write a single line?

What I did

The first thing I did was connect Claude → Slack. All my conversations with the client were already there. Instead of prompting the brief from scratch, Claude could just read the actual thread. The real words the client used, the back and forth, the things they kept emphasizing. That context matters more than people think.

Then we started brainstorming. And I mean really working through it, not “give me a plan” and copy-paste whatever comes back. Going deep on the questions that actually matter: what does this pipeline need to do at every step? Where does data move? What happens when something breaks? What are the edge cases nobody mentioned yet?

I burned through my entire Claude Pro session doing this. This is me waiting for the limit to reset:

It took about two to three days to map it all out properly. That might sound long for a planning phase. But here’s what happened during those sessions: I found gaps I had no idea existed. Things that were completely invisible to me until I started asking the right questions out loud, with Claude pushing back and helping me think. Issues I would have hit halfway through the build, at the worst possible moment, when changing them would have cost me hours.

We put everything into a structured architecture document in Notion as we went. Not a rough sketch. A proper reference for the whole system.

After that I needed to actually see it. So I used a Claude skill to generate diagrams of the full pipeline. Components laid out, data flows mapped, dependencies visible. There’s a big difference between understanding something in your head and being able to look at it on a page and spot what’s off.

Last thing before opening any tool: a todo list. Every workflow to build, every credential to request from the client, every API to connect. Exact steps, exact order. No pausing mid-build to figure out what comes next.

By the time I actually started building, the hard thinking was already done.

Here’s the end-to-end flow and how I currently structure the delivery phase of all projects I have:

The build went smooth. Not because everything went perfectly, but because when something came up I already knew how the system was supposed to work. The plan was there to come back to.

I used to think AI was mostly useful for coding, writing this, generating that. And it is.

But the best thing I got out of this project was the two days before I opened n8n. The brainstorming, the gaps I found early, the clarity I had before I touched anything. That alone saved me hours of mid-build pain. That’s what I’ll take into every complex project from here.

Lesson learned

The thing is that nobody shows you this part.

Every AI influencer on your YouTube feed is posting a 60-second clip of a workflow running. The automation fires, the data moves, the result appears. Clean. Impressive. Done in 10 minutes.

What they don’t show you is the two days before that.

The part where you actually have to think. Where you sit down with the problem, ask uncomfortable questions, and find the holes in your plan before they become holes in your build. Because that surely doesn’t make a good video.

“I stared at a Notion doc for two hours figuring out edge cases” doesn’t get 50,000 impressions. But it’s the reason the build went smooth.

AI gurus are selling you speed. Which might be real in some cases, especially for simpler builds. But speed without clarity just means you hit the wall faster.

So to make a long story short here’s one thing you can take out of this article:

Plan with AI before you build with AI.

Talk soon,

Ilya

Every AI Guru Is Lying to You

Not Maliciously. Just Conveniently.

• “Automate your entire business with AI.”

• “Build in a weekend.”

• “No code needed.”

• “While you sleep.”

Here’s the thing nobody says out loud: most of them are not actually running a business. They’re selling the idea of running one. The content is the product. And you are the customer.

So let me tell you what automation with AI actually looks like when you use it on real client work, with real deadlines, and real money on the line.

I’ve been doing this since 2025, dozens of projects across marketing, sales, ecommerce, and content. And the truth is more boring and more useful than anything you’ll see in a carousel post.

The Stack (And Why It’s Boring on Purpose)

I use Notion, n8n, and Claude. That’s it. Yep…that’s it!

Not a stack of trendy SaaS tools. Not a Zapier chain with 40 steps held together with hope. Just Notion, n8n, Claude, and a set of well-written instruction files with a folder on my computer.

Notion handles everything project and client related. Tasks, docs, call notes, proposals. It connects cleanly to most things you already use.

n8n is the automation layer. It’s where I connect APIs and build the actual workflows. I use it for three things mainly:

• Document Automation. Turning recorded sales and onboarding calls into beautiful summaries, proposals, or overviews automatically.

• Lead Gen + Qualification. Scraping, cleaning, and qualifying lead lists.

• Content Automation. Iterating on content drafts without starting from scratch every time.

Claude is where the thinking happens. But the reason it works well for me is something called skills.

Skills are exactly what they sound like. Each one is a set of instructions, a markdown file on my computer, that Claude reads before doing a specific job.

One skill tells it how to scrape a LinkedIn profile and qualify the lead. Another handles spreadsheet output. Another builds Notion proposals on the fly with proper formatting.

Before I had this, every task started from scratch.

• Explain the context

• Explain the format

• Explain the tone

• Explain what I want

Every single time. Like hiring a contractor and having to explain the paint color every time they move to a new wall.

Now I just say: use the proposal skill, here’s the topic from my last notion call with XYZ lead. It already knows the rest.

What this actually replaced

Let me be specific, because vague claims are how you end up buying another tool you don’t use.

Client case studies. Used to take me 2 hours minimum. Read through my notes, structure the story, write the draft, edit the draft. Now I pull the project data from my Notion workspace, tell Claude which project to use, and have a full draft in under 10 minutes. I spend 15 minutes editing. That’s it.

Lead qualification. I used to go through lead lists name by name. Open LinkedIn, read the profile, decide if it fits, move on. For a list of 50 people that’s a morning gone. Now I paste the list, have an ICP in my Claude instructions, and get a scored table back with reasoning. I spot-check a few rows, adjust if needed, and move on.

Client deliverables. Proposals, onboarding documents, project overviews. These used to mean opening last month’s version, stripping out the client-specific parts, and rebuilding. Now the skills know the structure. I give Claude the inputs. It produces the document given a specific template. I review it.

None of this is magic. It’s just documented process, handed to a model that can execute it without getting bored or losing context.

Where people go wrong

Now, the gurus will tell you the hard part is finding the right tool. That if you just get the right stack, the automation handles itself.

It doesn’t. You have to do 2 things at first:

• Understand your own workflow well enough to describe it as clearly as possible.

• Know how to combine every piece of the puzzle properly to get real results.

That’s when it starts to feel like magic.

Here’s my number 1 rule when it comes to automation:

Manual first, automate later.

Most people have never written down how they actually do their job. They just do it. And when you can’t explain what “good” looks like for a task, no AI in the world will produce it consistently.

The skill is only as useful as the instructions behind it.

This is the real work. Not prompt engineering. Not building agents. Not following whatever framework dropped on YouTube last Thursday.

It’s sitting down and doing a lifestyle audit: what does this task actually involve, what does a good output look like, and what would I need to tell a smart new hire to do it right?

Once you have that answer written down, you have a skill. And you can reuse it as many times as you want.

The honest version of “AI will grow your business”

It will. But not because the AI is impressive.

Because every hour you stop spending on repeatable tasks is an hour you can spend on the things that actually move the business forward. Conversations with clients. Scoping new work. Building relationships. Thinking.

The automation isn’t the growth. The time it frees up is.

That’s the version nobody’s selling you, because it doesn’t make for a clean marketing hook.

But it’s the one that actually works.