Stop re-training your AI every single day. Your edits are a system
After 15 corrections you close the chat — and throw away the most valuable thing: the path from bad output to good. Here's the prompt that turns edits into a compounding instruction.
01 — The pain
Recognize this daily ritual?
You open a new chat and explain, again:
- no corporate-speak;
- don't open with platitudes;
- too soft here;
- this example has nothing to do with my audience;
- I've asked ten times not to use the word "unique" 😅
After fifteen corrections the AI finally produces something decent. You take the text, close the chat — and tomorrow you start over. Same corporate-speak, same platitudes, same "unique."
It's not because "AI has no memory" or "the model is dumb." It's because you threw away the most valuable part of yesterday's session yourself.
02 — The shift
What's actually the most valuable thing in your chat?
Not the final text. The path from bad output to good.
Because inside your corrections lives an instruction: how you think, what you consider quality, which mistakes you notice, why you change direction, and by which criteria you accept the final version. Fifteen corrections are fifteen standards you articulated — standards no model ships with.
The final text. Used once — done: tomorrow the model writes corporate-speak again and you spend another half hour on the same explanations.
The system inside the edits: "I check headlines by this criterion," "that audience is wrong," "the word 'unique' is banned." Extracted once — works in every future task.
Your unconscious expertise already surfaced in the dialogue. It just needs to be pulled out — and one prompt does that.
03 — The cycle
How does the cycle that compounds your decisions work?
The whole system is four steps that close into a loop. Walk through them:
Run your session like always: set the task, correct, push to a decent result. Don't document anything along the way — the corrections stay in the dialogue, and that's your raw material.
Before closing the chat, send the prompt from the next section. The model will dissect the whole dialogue and extract the system from your edits: what didn't work for you, why the direction changed, which phrasings you approved, which mistakes must not repeat.
If your AI agent has access to a knowledge base — it updates the file itself. In a regular chat — ask for a ready file and save it to a project (ChatGPT Projects, Claude Project knowledge). One file, e.g. "Content_Rules.md".
New session — first have the AI read the rules file, then set the task. Run today's corrections through step 2 again: the file gets updated, duplicates and contradictions removed. Every session makes the next one shorter.
04 — The prompt
Which prompt extracts the system from a dialogue?
Here it is in full — send it at the end of every important working session:
Analyze our entire dialogue. Do not summarize the topic and do not save only the final result. Extract a working system from my corrections: what the original task was, what I corrected, what exactly didn't satisfy me, why the direction changed, which decisions I made, which phrasings I approved and which mistakes must never repeat. Turn this into a clear instruction for the next task of this kind. Merge the new conclusions with the existing rules, remove duplicates and contradictions, and update the file in my knowledge base.The key words are "do not save only the final result." Without that line, the model retells the outcome. With it, it reads the dialogue the way an editor reads revision history: what changed and why. The "why" is what becomes rules.
BANNED: "unique", "in today's world", opening with definitions
HEADLINES: must raise "what happens next?"; no promises
the text doesn't deliver
AUDIENCE: experts 30–45, tired of info-noise; never write
"for AI beginners"
TONE: human, some self-irony; soft ≠ vague
ENDINGS: always a practical step, never an "inspiring conclusion"05 — The effect
What happens after a month of this cycle?
Today you explained why that headline was weak. Tomorrow the model checks headlines by your criterion itself. Today you stopped it when it drifted to the wrong audience. Tomorrow the right audience is already written into the instruction.
And no, the file isn't carved in stone. Every new dialogue extends it: old rules get refined, contradictions get removed. Your unconscious expertise gradually turns into a working system.
That's where personal AI actually begins. Not when the model "magically remembers you." But when you stop losing your own decisions and start accumulating them in a knowledge base.
06 — The start
Where do you start today?
No setup needed in advance. Today's working session is already material:
- Run a normal working session with corrections (or open yesterday's long chat)
- Send the analysis prompt from section 04
- Save the resulting rules file to a project (ChatGPT Projects / Claude Project)
- Tomorrow, open the session with: "read the rules file first"
- In a week, run the analysis again — the file updates and de-duplicates
Every correction is either a one-off waste of time or a brick in a system. The difference is one prompt at the end of the session. Stop throwing away the path from bad to good — that's where your personal AI actually lives.
FAQ
How is this different from ChatGPT's built-in memory?
Built-in memory stores facts about you at its own discretion — you don't control what it kept. A rules file is your conscious decisions extracted from real corrections: you can read it, edit it, clean contradictions, and carry it between models and tools.
Where should the rules file live?
Wherever the model will read it at session start: ChatGPT Projects, Claude Project knowledge, or your agent's knowledge base — then it updates the file itself. In a plain chat, just ask for the file and paste it at the start of each new session.
How often should I run the analysis prompt?
At the end of every substantial working session — the ones with lots of corrections. It takes a couple of minutes. The prompt itself merges new conclusions with existing rules and removes duplicates, so the file doesn't sprawl into a junk drawer.
Does a rules file replace a knowledge base and context?
No — it's the ground floor. The rules file covers HOW you work: criteria, bans, tone. A knowledge base covers WHAT you know: cases, numbers, materials. Together they give the model both your standards and your substance — start with the rules; it's the fastest win.
