Lesson 3 · 15 min
When the AI already knows who you are, what the company rules are and what went wrong last time, your request fits on one line. You don't build this base in every conversation: it lives in a place the AI reads on its own before it starts. It's what separates repeating everything every time from having someone who already knows the work.
Start the lesson
Goal: build the context base the agent reads before any request, with rules and a lesson from a real mistake.
Four sources the agent reads before your request. Tap or click each one.
Scroll the page: each step below lights up its part of the diagram.
One file per project (AGENTS.md in Codex, CLAUDE.md in Claude Code, the project instructions in ChatGPT or Claude) with who you are, your audience, your tone, your tools and your boundaries. The agent reads it before it starts.
Every time something goes wrong, add a new line at the top of the list: what happened and the rule that keeps it from happening again. In my case (Nei, from INEMA), it works like this: the same mistake doesn't come back, because the fix becomes a written rule.
What you've already decided with the AI shouldn't be asked again. In the chat, turn on ChatGPT or Claude memory in the settings and keep the topic inside the same project. In the terminal, save the session logs in a folder the agent can search.
The real documents and spreadsheets: in the chat, the files uploaded to the project; in the terminal, the folder or repository the agent opens. It uses the original instead of your memory-based summary.
With the four sources in place, the request gets short, and the agent understands where it came from and why. You don't build the context from scratch every time.
Rules, lessons, history and files reach the agent before the request. Each mistake becomes a lesson, and the next task already starts better.
Good context has a home of its own: it lives in a file or folder, not in the text of the request.
Start small. Tap or click each step.
In the chat: create a project (ChatGPT or Claude) or a Gem (Gemini), write the rules in its instructions and upload the reference files.
In the terminal, with Claude Code or Codex (agents that work in a folder on your computer), the rules go in a file in the folder called AGENTS.md or CLAUDE.md. That's just the name of the rules file.
A project called "School communication" with its own instructions.
Start with five rules: who you are, who the work is for, the tone, the tool and one boundary.
A good rule is checkable: you can look at the result and say whether it was followed.
Every announcement states the date the change starts.
When the AI gets something wrong, don't just fix that one answer: write the lesson at the top of the list, with the date.
One line: what happened and the rule that prevents it.
Announcement went out without the start date: every change states the day it starts. (10/24)
Open a new chat, make a short request, and see whether the rule and the lesson were followed without you reminding it.
If they weren't, the rule is buried or vague: move it to the top and word it so it can be checked.
Write the notice for the holiday party.
Replace the text between < > with your own.
# Who I am <your name, what you do, who you work for> # Who the work is for <the audience, in one sentence> # Rules - Write in English, short sentences. - <checkable rule 2> - <checkable rule 3> - Before you finish, tell me what changed and how I can check it. - Boundary: <what must never happen> # Where the rest is - References: the files uploaded to the project (in the terminal, the references/ folder) # Lessons (newest on top) - <date> · <what went wrong> → <the rule that prevents it>
You got this wrong: <what happened>. Before fixing it, write the lesson in ONE line, in this format: <date> · <what went wrong> → <the rule that prevents it>. Then redo the work following the new rule and show me both.
Read the project's rules and lessons before you start. If anything I ask goes against a rule, tell me before you do it.
Pick a subject where you use AI every week.
Create the project (in the chat) or the folder with AGENTS.md (in the terminal) and write the five rules from the template.
Done when: the file or the project instructions have five rules, and each one can be checked by looking at the result.
Ask for a task, find a mistake (there always is one), use the "when the AI gets it wrong" request, and paste the lesson into the project instructions (in the terminal, it saves it in AGENTS.md).
Done when: the Lessons section has a line with the date, the mistake, and the rule that prevents it.
Open a new chat and make a one-line request of the same kind.
Done when: the result followed the rule and the lesson without you repeating either one in the request.
The explanation appears after your answer.
Question 1 of 3
The AI forgot the date in the announcement. What can you do so the mistake doesn't come back?skill
Question 2 of 3
Which of these is a checkable rule?
Question 3 of 3
Why should the context have a "home of its own"?
Choose an answer. The explanation appears after you try.