Iterate: correct, ask for alternatives, ask for a critique
This lesson is still being tested: it improves with learners' feedback. If something is unclear, it is the lesson that needs fixing, not you.
By the end of this lesson, you will be able to
- correct one specific point without losing what was good;
- ask for variants and choose the best one;
- have the AI review an answer, using your criteria;
- know when to start over in a new conversation, with the essentials.
Why it matters to you
A first answer is rarely perfect, and that is normal: it is a draft. Many people stop there, disappointed, or start everything over with different wording. The right reflex is to iterate: keep what works, correct the rest, in two or three messages.
Before you start
| You need | Your assistant, free account |
| Cost | €0 |
| Duration | 35 minutes |
| Reminder | The email in the example is made up; for your real messages, nothing sensitive |
Understand it in 3 minutes
Four steps to improve an answer:
| Step | When to use it | Example |
|---|---|---|
| Correct one point | It’s almost right | “Keep everything, but replace the 2nd paragraph: too formal.” |
| Ask for variants | You can’t decide between several styles | “Give me 3 subject lines: a neutral one, a firm one, a very short one.” |
| Ask for a critique | Before using the result | “Reread your answer using my criteria, and tell me honestly what isn’t met.” |
| Start over cleanly | The conversation is getting muddled | “Summarize in 5 lines what we decided”, then a new conversation with that summary |
Say what is wrong, and what is right. “Redo it” or “it’s not good” don’t say what to change: the AI changes everything, including what worked. “Keep the structure, change the tone” tells it exactly where to work.
The AI’s review helps, but does not replace yours. Asking the AI to critique its answer can sometimes point out something missing. But an AI can be confidently wrong. OpenAI acknowledges this for ChatGPT: the model can sound very sure of itself, even when it is wrong. That is no reason to distrust everything. You always reread it yourself, but not everything in the same way: tone and form can be judged at a glance; facts (figures, names, dates, rules) get checked when a mistake would cost you (lesson 4).
A muddled conversation is hard to repair. If the AI repeats a mistake you already corrected, or mixes up two versions, don’t keep pushing: have it summarize what has been decided, and start a new conversation with that summary.
Do it together
The case is made up: your heating has been broken for 5 days, and you are writing to your landlord.
1Step 1: The first draft
Write an email to my landlord asking them to get the heating repaired: it has been broken for 5 days.
Form: under 10 lines, polite.
It worked if the email gives a clear deadline for someone to come and fix it.
Read the answer: it is a draft. Notice what you like, and what is off.
2Step 2: Correct one point, without losing everything
Keep the structure, but make the tone firmer, and add that I already called twice without a reply.
Check that the rest has not changed. If the AI rewrote everything, say so: “You changed the ending, put it back the way it was.”
3Step 3: Variants
Give me 3 subject lines for this email: a neutral one, a firm one, a very short one. Tell me which one you would choose, and why.
Choose your own. You can also combine two.
4Step 4: Ask for a critique
Reread the email as if you were the landlord: what could annoy them, or lead them to wait? Then fix those points.
Your check: reread the final version yourself. If the email cites a law, an article or a legal deadline that you did not provide, it is information to check (lesson 4), or to remove.
5Step 5: Start over cleanly
Summarize in 5 lines what we decided for this email, so that I can start again in a new conversation.
Reread the summary: are all the decisions and constraints in it? Complete it if needed, and also keep the last version of the email that suited you. Open a new conversation, paste everything, and ask for the final version. You now know how to start over in a fresh conversation with the essentials, checked by you.
Your turn
A real text you need this week: a message, an ad, a reply to an email, a description.
- Get a first draft with a complete request (lesson 1).
- Use each of the first three steps at least once: correct one point, ask for variants, ask for a critique.
- Reread the final version yourself, and note how many messages it took.
Check your work
It worked if:
- you have a version you can really use, in 5 messages at most after the first draft;
- each correction said what to change, and what to keep;
- you reread the final version yourself, without stopping at the AI’s critique;
- nothing in the text comes from information that you neither provided nor checked.
Common mistakes:
- “Redo it” without saying what to change: the AI rewrites everything, including the good parts;
- losing a good version: copy it before asking for variants;
- taking the AI’s critique for a check: it is a help, not proof;
- piling up fifteen corrections in a muddled conversation, instead of starting over with a summary.
Keep it and pick it up later
- Add to “My requests” the correction that helped you most, for example “Keep everything, but…”.
- Keep the summary request from step 5: it will help you every time a conversation gets too long.
Going further
optional
- Try role-playing for the critique: “Reread this text the way [the person who will read it] would.” A recruiter, a customer or a neighbor do not notice the same things.
New words
4 words explained
| Word | What it means |
|---|---|
| Iterate | Improve a result in small steps, instead of starting over |
| First draft | The first version of a text, which you then improve |
| Variant | Another version of the same content, in a different style or length |
| Targeted correction | A request that says precisely what to change, and what to keep |
Check your understanding
3 questions, no grade. An explanation appears after each answer.
What's next?
Lesson 4 · Check: sources, cross-checking, limits of web search: know where a piece of information comes from, and whether you can believe it.
Module 2:
✏️ Do the exercisesMy space: your progress, your exercises and your last lesson.
Follow this lesson with Claude Code
With the course companion installed, open Claude Code in your working folder and type:
/formation-ia:suivre m02-l03 enIt reads the up-to-date lesson on this site and guides you step by step, without changing anything without your consent. Install the companion. Claude Code is not included in the free plan: an eligible Claude subscription, or a Console account billed by usage for the terminal.
Sources and checks
1 sources
Date, version, sources Written on 24/10/2026. The method in this lesson works with any assistant. The sentence about ChatGPT’s confidence comes from OpenAI’s help page (“Does ChatGPT tell the truth?”, help.openai.com), read on 24/10/2026.
Sources checked: