How do I refine a prompt that gives a disappointing answer?

Do not rewrite the whole thing. Write one sentence on what exactly is wrong (too long, tone too vague, a number missing), then ask for that correction only. If the same gap keeps coming back, the starting task was incomplete.

What is prompt iteration?

Prompt iteration is correcting an AI's answer through short, targeted follow-up messages instead of starting over. Each message names one gap between what you got and what you wanted, so every round is easy to read and to undo.

Why change only one thing per message?

If you change the tone, the length and the example in one message, you cannot tell which of the three moved the answer. One change per message makes each round readable and lets you go back to the best version.

How many rounds of iteration should I expect?

Two or three targeted corrections are usually enough. Beyond that, reread the starting task: a missing success test or missing context is faster to fix at the source than in five patches.

TrainingPublished 6 October 20263 min

How to Refine a Prompt: Set the Task, Then Iterate

By Alexandre Saint-Jean

Most people judge an AI assistant on its first answer. That is a mistake: the first answer is rarely the right one, and it is not meant to be. What makes the difference is how you set the task, then how you correct it.

What does a well-posed task look like?

A well-posed task has three parts: a precise action verb, an identifiable deliverable and a success test. "Help me with my email" says none of this. "Write a follow-up of 80 words maximum, offering two time slots, in a direct but polite tone" says all three.

Anthropic's guide offers a simple check: show your prompt to a colleague who knows nothing about the subject and ask them to carry it out. If they would hesitate, the AI will too. The same guide recommends numbered or bulleted steps when the order or completeness of the steps matters.

Why is the first answer rarely the right one?

Because even a clear instruction leaves choices open: the angle, the level of detail, the register. The assistant makes those choices by default, and they are not always yours. OpenAI's guide puts it in one line: getting the output you want is a mix of art and science, which is why it recommends, for a prompt used in production, building tests that measure its behavior rather than trusting a hunch.

Google's guide for Gemini notes the same thing: using different words or phrasing, even to say the same thing, often yields a different response. Iterating is normal work, not a sign the first attempt failed.

The trap is to reply "make it better" or "no, not like that". The assistant then has to guess what is wrong, and it often guesses wrong.

How do you iterate: name the gap, change one thing at a time

A good correction names the gap. "Keep the second paragraph, shorten the first, and replace the word opportunity with a number" beats ten rounds of "better".

The second rule is to change one point per message. If you ask to shorten, change the tone and add an example at once, you cannot tell which of the three moved the answer. With one point per message, every attempt teaches you something and you can easily return to the version that worked best.

A step-by-step example: a sales follow-up

Task: "Write a follow-up of 80 words maximum that offers two time slots, in a direct but polite tone."
Gap 1: "Keep the text, but open with the benefit for the client, not for us."
Gap 2: "Replace the last sentence with a closed question about the first slot."

Three messages, three readable changes, no going back.

When should you stop iterating?

Two signals. If the same correction comes back in every conversation, it no longer belongs in iteration: it should move into your account or project context, as covered in our guide to account and project context. And if you cannot describe what you want, show it: giving three examples instead of rules often solves the problem in one go.

This progression (prompt, context, task, examples) is the one our free AI newsletter follows, one idea per week.

Frequently asked questions

How do I refine a prompt that gives a disappointing answer?: Do not rewrite the whole thing. Write one sentence on what exactly is wrong (too long, tone too vague, a number missing), then ask for that correction only. If the same gap keeps coming back, the starting task was incomplete.
What is prompt iteration?: Prompt iteration is correcting an AI's answer through short, targeted follow-up messages instead of starting over. Each message names one gap between what you got and what you wanted, so every round is easy to read and to undo.
Why change only one thing per message?: If you change the tone, the length and the example in one message, you cannot tell which of the three moved the answer. One change per message makes each round readable and lets you go back to the best version.
How many rounds of iteration should I expect?: Two or three targeted corrections are usually enough. Beyond that, reread the starting task: a missing success test or missing context is faster to fix at the source than in five patches.

Sources

Go further

Glossary: what is a prompt