Ask your own history what works

I thought my problem on LinkedIn was the algorithm. It was me.

Last week I pulled every post I have written since 2018. 298 of them. I gave them to our company brain and asked one question: what works, and what doesn’t?

The answer was a little embarrassing. Since January 2025, my long Arabic posts with one image averaged 6.8 comments each. My short English posts averaged 1.1. And in October I had been posting almost nothing but the second kind: a quote repost, one English line, a newsletter link.

I was writing for the algorithm in my head, not for the people who actually read me.

The lesson is bigger than LinkedIn. Anything you have done for years has a written history: emails, proposals, posts, reports. That history knows things about you that you have stopped noticing. This is the method, step by step, with the prompts. It needs no code and about an afternoon.

What the numbers said

Here is what came back, measured on the 195 posts since 1 January 2025. Comments, not likes, because a comment is the start of a conversation.

The day mattered too. Wednesday averaged 7.5 comments a post. Monday 3.1, Thursday 2.8. I had been posting mostly on Mondays.

None of this was a secret. Every number was sitting on my own profile. I had just never put them side by side.

Step 1: pick one kind of work and one number

One kind only. And one number that says whether it worked.

If you cannot name the number, stop here. A history without a result is just an archive.

Step 2: put the last 50 in one sheet

One row each: the date, the text, the number, and the things you suspect matter. For posts that was language, length, format (photo, video, carousel) and the day. Fifty is the floor. More is better; I used 298.

For LinkedIn, Settings, then Data privacy, then Get a copy of your data gives you the text of your posts but no numbers. The numbers come from each post’s analytics, or from a scraping tool. I used Apify. For emails, export the folder. For proposals, the file name and a won or lost column is enough to start.

Leave out anything that holds other people’s private data. You do not need it to find a pattern.

Step 3: ask, from the file only

Give the sheet to the AI and use these in order. The one rule that matters: it answers from the file, not from what it thinks good writing looks like.

Here are my last [N] [posts]. Each row has the text and one result
number. Split them by language, length, format and day. Show the
average result and the row count for each group. Use only this file.
No opinions.

Compare my top 10 and bottom 10 by result. What do the top 10 share
that the bottom 10 don't? Quote the rows behind each point.

Which of these patterns rest on fewer than 10 rows? Mark them as weak
signals, not rules.

Write at most three rules from this. One line each, with the number
behind it.

The third prompt is the one people skip. My three comment-gate posts averaged 33.7 comments. That is the best number in the whole sheet, and it is three posts. A strong signal, not proof.

Step 4: write the rules in a file the agent reads

A rule in your head lasts about two weeks. Then the old habit wins.

So the rules went into one file, and the agent that helps me write reads it every time before it starts. Mine looks like this:

# rules.md · what works, from my own history
1. Write in Egyptian Arabic, 900+ characters.   (6.8 vs 1.1 comments)
2. One image, never a carousel.                 (6.4 vs 1.8)
3. No quote reposts.                            (0.7 comments)
4. The strongest idea goes on Wednesday.        (7.5 vs 3.1 on Monday)
Updated: 11 Oct 2026. Source: 298 posts, Jul 2018 to Oct 2026.

Where it lives depends on your tool. In Claude, put it in the Project instructions or the folder the agent reads, and start every prompt with one line:

Read rules.md before you start.

In ChatGPT, the same file goes into a Project. With nothing else, paste it above every draft.

Step 5: close the loop every week

A rules file you never check turns into a style guide: what you believed on the day you wrote it.

From this month, one of our agents does this every Monday: it reads the week’s posts, compares them with the rules and tells me what followed them and what did not. If you do it by hand, this is the prompt:

Here are this week's [posts] and their results. Which ones followed
rules.md and which didn't? Did the rules hold? Suggest one change to
the file, with the number behind it.

The part that made it stick: the cockpit

The numbers alone would not have changed anything. I know myself. Two weeks later I would have been back to the old habit.

What changed it is that the result turned into daily work, in one place. We run Athr21 from a screen we call the cockpit. I open it every morning on my phone. It holds everything that is waiting on me: a post to comment on, a message to answer, a person I need to call. The agents put those things there. I decide.

So the rule from my history is not a note I have to remember. The agent that drafts reads it. The agent that reads my posts on Monday checks it. And the cockpit is where I see what is left for me to do. A step you have to remember is a step you skip.

Honest limits

A small sample is a signal, not a rule. Under ten rows, do not build on it.

Match every result to the right piece by its text, never by its date. We got this wrong once: a session matched a post by date, picked the wrong one and wrote its numbers into three files before anyone checked. Now nothing is written until the text matches.

Likes on posts that tag people are lifted by the people tagged. Older posts reached a smaller audience. And your history only knows what you already tried. It cannot tell you about the format you never used.

I even broke rule 4 on purpose. The post that points to this article went out on a Thursday, as a test. Next week I will know whether Thursday deserves its 2.8.

Do it yourself

If you already work inside an AI workspace, paste this and start:

I want to learn what actually works in my own work, from my own
history. Work through these stages in order and wait for my review
between each:

1. PICK. Ask me which kind of work to study (posts, proposals,
   follow-up emails, reports) and the one number that says it worked.
2. SHEET. Help me build one sheet of my last 50 or more: date, text,
   result, and the traits I suspect matter. Leave out other people's
   private data.
3. GROUPS. Split the rows by each trait and show the average result
   and the row count per group. Use only the sheet. No opinions.
4. TOP VS BOTTOM. Compare the top 10 and bottom 10. Quote the rows
   behind every point. Mark any pattern on fewer than 10 rows as weak.
5. RULES. Write at most three rules, one line each, with the number
   behind each, into rules.md. Read rules.md before every draft.
6. LOOP. Every week, check the new results against rules.md and
   suggest one change, with the number behind it.

Rules: never invent a number. Match results to items by their text,
never by date. Never publish anything on my behalf.

The deeper point needs no AI. Your work already has a memory. It is in your sent folder, your proposals and your old posts. Asking it what worked is the cheapest research you will ever do. That is the same habit The Brain builds inside a company: decisions and results written down where the whole team, and its agents, can read them.