Iterate: critique, refine and retry
Your first answer is a draft. Learn a practical loop for improving it: give specific feedback, ask the model to critique itself, change one thing at a time and know when to start a fresh chat instead of pushing on.
Nobody gets their best result from the first prompt, not even people who write prompts for a living. The skill is in the next three minutes: reading the answer critically and steering it. This tutorial gives you a loop for that, using a cover-letter email for an internship as the running example.
You need a free chat tool (ChatGPT, Gemini, Claude or Copilot) and about 25 minutes. You should have read *Anatomy of a good prompt* and *Be specific*.
The iteration loop
Four moves, repeated
- 1Run: send a clear first prompt and read the answer fully.
- 2Diagnose: decide what exactly is wrong (missing context, wrong tone, wrong format, wrong facts, too long).
- 3Fix the cause: change the prompt, or send a targeted follow-up.
- 4Compare: check whether the new answer is actually better than the last, then stop when it's good enough.
The most important move is diagnose. "Make it better" is as vague as the first prompt was. Decide what is wrong first: that tells you what to add.
Round 1: a first draft
Write a short email to a hiring manager applying for a summer internship in web development.
About me: I am a second-year BCA student in Pune. I built a React website for my uncle's tailoring shop and I know HTML, CSS, JavaScript and basic Git. I have no work experience.
Constraints: under 120 words, polite and direct, no clichés like "I am writing to express my keen interest".What to expect: You will probably get a decent but generic email. Read it and write down what you would change before you continue.
Round 2: give specific feedback
Vague feedback vs useful feedback
| Vague | Specific |
|---|---|
| Make it better | Replace the opening line with one concrete fact about my project |
| Too long | Cut to 90 words; remove the last paragraph |
| I don't like the tone | Make it less formal, like a message to a senior colleague, not a letter to a judge |
| It sounds fake | Remove claims I haven't made ("passionate", "quick learner") and mention the tailoring site instead |
| Not what I wanted | Keep paragraph 2 exactly as it is. Change only paragraph 1 |
Keep the second paragraph exactly as it is. Change the opening so it starts with one concrete fact about the React site I built for the tailoring shop. Remove any claim about my personality (such as "passionate" or "hardworking"). Keep it under 120 words.What to expect: Check that paragraph 2 really did not change, and that the opening now contains a fact from your message. If the model rewrote everything, say so and ask it to restore the original paragraph 2.
Say what to keep
Models often rewrite everything when you ask for one change. Telling them what to keep unchanged ("keep paragraph 2 exactly as it is") protects the parts you like.
Round 3: ask the model to critique
If you can't tell what is wrong, ask the model to be the critic. The trick is to give it a role, criteria and a request to find problems, not to praise. Without that, models tend to say "this is a great email".
Act as a tough but fair reviewer. Review the [email / essay / code / plan] below against these criteria: [e.g. clear in the first line; specific evidence; right length; right tone for a hiring manager].
For each criterion, give a score out of 5 and one specific problem with a quote from the text. Then list the 3 changes that would improve it most. Do not rewrite it yet.
[email / essay / code / plan]:
---
[paste your text]
---What to expect: Good critiques quote the exact weak sentence and explain why it is weak. If the critique is only praise, ask: "Assume a stricter reviewer. What would they reject this for?"
Rewrite the text applying changes 1 and 3 from your critique only. Leave everything else as it is. Then list what you changed in one line each.What to expect: Choosing which suggestions to apply is your decision: some suggestions are taste, and some may conflict with facts about you.
Self-critique has limits
A model reviewing its own writing shares its own blind spots. It can miss invented facts, repeat the same style problems and make confident but wrong suggestions. Treat critique as a second pair of eyes, not a judge. You decide, and you check facts.
Change one thing at a time
If you change the role, the length, the tone and the examples all at once and the answer improves, you don't know why. Models also have randomness: the same prompt can give different answers, so one good or bad answer proves little. To find what actually helps:
A small experiment
- 1Run your current prompt two or three times in new chats and note the typical result.
- 2Change exactly one thing (for example, add a sample sentence in your style).
- 3Run it two or three times again.
- 4Keep the change only if the results are better most of the time.
- 5Save the prompt that works so you can reuse it (see *Prompt templates and your personal prompt library*).
Retry, regenerate or restart?
What to do when it's not working
| Situation | Best move |
|---|---|
| Close but one part is off | Targeted follow-up in the same chat |
| Style is fine, wording varies from run to run | Regenerate once or twice and pick the best |
| The answer misunderstood the task | Edit your original prompt and rerun, don't argue in the chat |
| After 4-5 rounds of fixes it got worse or contradicts itself | Start a new chat with a cleaned-up final prompt |
| It keeps inventing facts | Give it the facts and add "use only these"; verify the result yourself |
Long chats pile up old instructions, rejected drafts and corrections, and models can lose track of them. A fresh chat with one clean prompt that includes what you learned is often better than a seventh round of "no, not like that".
We have iterated on [the task] in this conversation. Write one single clean prompt that I can use in a new chat to get the final version directly. Include the context, constraints and format that we ended up with, and the corrections I made (for example [what you fixed]). Return only the prompt.What to expect: Read the prompt it gives you and fix anything it got wrong, then try it in a new chat. If the new chat gives a good result at once, save the prompt.
Know when to stop
Good enough?
- Every fact in the answer is true and from you or a source you checked.
- It meets each constraint you set (length, format, must-haves).
- A real reader would understand it on one read.
- Remaining issues are small enough to fix by hand faster than by prompting.
Iterating forever is its own trap. When the last two rounds only changed words and not substance, finish it by hand.
Try it: Run the loop on your own draft
Take a real piece of writing: a statement of purpose paragraph, a LinkedIn About section or a README. Run one full loop: first prompt, diagnose in writing (what exactly is wrong, in 2-3 points), a targeted follow-up, the critique prompt, apply two suggestions, then use the summarise prompt and test it in a new chat. Compare the first and last versions side by side.
Key takeaways
- The first answer is a draft; treat prompting as a loop of run, diagnose, fix and compare.
- Diagnose before you ask: say exactly what is wrong, and say what to keep.
- Ask the model to critique with a role, criteria and quotes; don't rely on it to judge itself.
- Change one thing at a time and run more than once, because outputs vary.
- Edit the original prompt when the task was misunderstood; start a new chat when a long thread gets messy.
- Stop when it's accurate and meets your constraints; finish small edits by hand.
Check yourself
Finishing the quiz marks the lesson as done.
1. Your AI email is too formal. Which follow-up is most likely to fix it?
2. You ask an AI "Is my essay good?" and it replies with only praise. What improves the next try?
3. Why change only one thing at a time when testing a prompt?
4. After five rounds of corrections, the chat's answers contradict earlier instructions. What is the best move?
5. What is a limit of asking a model to critique its own answer?
Saved on this device. Create a free account to keep it on every device.
Practise this with Sam
A real mock interview, out loud, with an honest scorecard. Free to start.