Asking Copilot to improve your prompt

1The three demands of a useful meta-prompt

Since beginner level, you have structured your prompts using four pillars: goal, context, expectations, and sources. The meta-prompt delegates this formatting discipline to the model itself for greater precision.

Research by Zhou demonstrates that model-optimized instructions consistently match or outperform human formulations. Major AI platforms incorporate this logic: your initial prompt represents a draft to refine methodically.

A solid meta-prompt states three clear requirements. First, a structured rewrite by named blocks: role, goal, context, sources, format, constraints, and verification guardrails. Second, a justification per change, in one sentence, identifying the specific defect fixed.

Third, a mandatory list of discovery questions regarding parameters the AI cannot guess: target audience, volume, and ambiguity handling. Always include the negative constraint: "do not answer the prompt itself", otherwise Copilot executes the raw command instead of refining it.

The limitation is well documented. Huang and co-authors showed that models struggle to self-correct without external grounding: a rewrite is unproven until tested on real data. Models frequently inject flattering roles, vague adjectives, and unenforceable rules. You remain the decision maker.

Tap to enlarge

2A sorting prompt rewritten block by block

A quality manager wants to classify the quarter's customer complaints.

Weak prompt, in the same conversation:

Improve this prompt: "Classify these complaints by topic."

Copilot returns a verbose prompt with a shallow "customer satisfaction expert" persona, "punchy" tone instructions, and five arbitrary categories.

Strong prompt: Turn 1 (diagnostic meta-prompt and block structuring):

Here is a prompt I use in Copilot Chat. Rewrite it according to the blocks Role, Goal, Context, Sources, Format, Constraints, Verification.
For each modification, 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 the modular rewrite, explains the need for guardrails, and asks four essential questions: what are the fixed categories, which column contains the verbatim text, and how to classify ambiguous complaints.

Strong prompt: Turn 2 (arbitrating suggestions and running execution):

I accept the block structure, table format, and uncertainty rule. I reject the expert role you added, which adds no value for factual sorting.
Here are my answers: 4 closed categories (Billing, Shipping, Product Quality, Other); text in column C; write "To qualify" if a complaint bridges two topics.
Apply this final prompt to this first batch of 10 complaints: [pasted data].

Copilot executes the streamlined prompt and classifies the sample without hallucination.

What changes: the meta-prompt exposes hidden ambiguities. Clarification questions elicit real business rules. Explicit arbitration prevents the model from injecting unwanted fluff.

3Have a prompt you use often rewritten

Take a prompt you use every week, or start from this one by clicking the composer's "Load exercise" button:

"Improve this prompt: Write the weekly project progress update from my notes."

To immediately test the prompt Copilot returns without searching for dummy files, here is a synthetic set of engineering site notes (fictional company NovaTech Ltd):

[NovaTech Ltd - Automation Project Site Notes (Exercise Data)]
Robotic Deployment Project:
- Milestone 1 (power enclosure cabling): completed 2 days ahead of schedule by tech crew.
- Milestone 2 (optical sensor shipment): supplier confirmed delivery delay of 5 business days.
- Executive trade-off: hold load test date on 28 September 2026 by mobilizing a temporary test bench.
- Priority action: Alice to follow up with freight carrier before Friday 5pm for customs clearance docket.

Write the meta-prompt that requests a rewrite by blocks, justification per change, and missing discovery questions (arXiv:2211.01910). Run it in Copilot Chat, answer the questions, then test the rewritten prompt on the notes above.

Reject at least one superfluous change with a written reason (e.g. a grandiose role like "World-renowned Agile Transformation Champion"). Self-assessment grid: (a) the rewrite is organized by named blocks; (b) each change is justified; (c) the final prompt was evaluated on the NovaTech notes and differs from Copilot's draft proposal on at least one point.

Open the prompt composer