Lesson 20 · Expert · 20 min
Asking Copilot to improve your prompt
The meta-prompt: have Copilot Chat rewrite your prompt block by block, with a reason for every change, then decide what you keep.
- Goal
- You will be able to have a prompt improved by Copilot Chat, require a justification for every change, and keep only the changes that serve your result.
- Skills
- Steer
Your first attempt, unaided
Paste into Copilot Chat a prompt you find average and simply ask it to improve it. Compare its version with yours, line by line, and look for what it added without telling you.
A meta-prompt is a prompt about another prompt. You paste your working prompt and ask Copilot Chat to rewrite it by blocks, explain each change and list what it lacks to do well. You remain the referee: you test the proposed version and keep only the changes that really improve the output.
1The three demands of a useful meta-prompt
Since beginner level, you have written your prompts with four ingredients: goal, context, expectations, source. The meta-prompt hands that structuring work to the model itself. Research showed it early: Zhou and his co-authors had a language model generate instructions, and those instructions matched or beat those written by humans on several task sets. Google Workspace recommends the "Improve this prompt" trick, and Microsoft reminds us that the first answer is a draft to be reworked.
A good meta-prompt makes three requests. First, a rewrite by named blocks: role, goal, context, sources, format, constraints, verification. Next, a justification per change, in one sentence, naming the problem fixed. Finally, a list of questions about what Copilot cannot guess: the audience, the volume, the rule in case of doubt. Add "do not answer the prompt itself", otherwise Copilot carries out the task instead of improving the request.
The limit is known. Huang and his co-authors showed that models correct themselves poorly without external feedback: a rewrite is not an improvement until it has been tested. The model readily adds a flattering role, adjectives ("professional", "punchy") and instructions it will not be able to follow. So you have to decide.
What Copilot often adds Keep it if Reject it if A role ("expert in...") It changes the way of answering (lawyer, trainer) It only flatters the model Tone adjectives They are translated into checkable criteria They replace an example or a figure A verification step It refers to a source provided It asks the model to "make sure" without a source 
Diagram "Ask Copilot to improve your prompt"Diagram generated by AI and reviewed 2A sorting prompt rewritten block by block
A quality manager wants to classify the quarter's complaints.
Weak prompt.
Improve this prompt: "Classify these complaints by topic."Copilot returns a longer prompt, with a "customer satisfaction expert" role, the instruction to be "structured and professional" and five invented categories. Nothing shows what changed or why, and the categories are not those of the department.
Strong prompt.
Here is a prompt I use in Copilot Chat. Rewrite it according to the blocks Role, Goal, Context, Sources, Format, Constraints, Verification. For each change, explain in one sentence the problem it fixes. End with the list of questions to ask me to complete what you cannot guess. Do not answer the prompt itself. Prompt: "Classify these complaints by topic."Copilot produces a rewrite by blocks, six justified changes (missing audience, undefined format, no rule for ambiguous reasons, unnamed source...) and four questions: which closed categories, which column holds the reason, what to do in case of doubt, what volume to expect.
What changes. The meta-prompt imposes a structure, so you see at a glance what was missing. Every change carries its reason, so you can accept or reject it knowingly. The questions bring you back to what only you know, instead of letting Copilot invent it.
3Have a prompt you use often rewritten
Take a prompt you use every week, or start from this one:
"Write the weekly project progress update from my notes."
Write the meta-prompt that asks for the rewrite by blocks, the justification per change and the missing questions. Run it in Copilot Chat, answer the questions, then test the rewritten prompt on real anonymised notes. The composer below helps you phrase the request field by field; you can also write it directly.
Reject at least one change with a written reason. Self-assessment grid: (a) the rewrite is organised by named blocks; (b) every change is justified; (c) the prompt kept was tested and differs from the proposal on at least one point.
4The meta-prompt that runs off and invents for you
A product manager types: "Improve this prompt: write the announcement for the new version of our app." Copilot returns a rewritten prompt and immediately follows up with the announcement itself. Won over by the result, he never reads the intermediate prompt. Yet that prompt made decisions for him: an announcement "aimed at key accounts", "to be published on LinkedIn", "highlighting the price cut", three invented assumptions, one of them false, since there is no price cut. What he should have seen: a meta-prompt that executes straight away turns every gap in the original prompt into an invention, and the polished output hides it. Correction: "Rewrite my prompt using the blocks Goal, Context, Sources, Format, Constraints. For every piece of information you had to assume, do not invent it: ask me the question. Wait for my answers before executing anything." Then reread the rewritten prompt line by line before running it, and strike out everything you never said. Rule to remember: a rewritten prompt is validated before it runs, never after.
5Quiz
Three questions, instant feedback. Each option comes with an explanation.
6Proof of mastery
Paste your starting prompt, the meta-prompt sent, Copilot Chat's proposal, then the prompt you finally kept, with the list of rejected changes and their reason, and an excerpt of the output obtained with the version kept.
Criteria
What you wrote at the start of the lesson
Going further
The meta-prompt reuses the blocks seen in the four ingredients and prepares the next lesson, evaluating a prompt on a set of ten cases, which gives a score to the version kept. The glossary term iteration sums up the loop.
Frequently asked questions
Does the meta-prompt replace writing the prompt?
No. It always starts from a prompt you wrote and a need you know. Copilot Chat knows neither who you work for nor what a good result is in your department. It structures, it questions, it proposes; you decide and you test.
Can Copilot improve a prompt without knowing my job?
It fills what it does not know with plausible generalities. That is why the meta-prompt asks for questions rather than assumptions: the list of missing information is often worth more than the rewrite itself.
Do you need a particular Copilot Chat mode?
The "Think deeper" mode in the model selector suits this analysis work well, because it plans and checks before answering. According to the documentation in force in September 2026, this mode is accessible without a paid licence, under standard access subject to service capacity.