4.4. Refining Prompts Based on Responses
💡 First Principle: A first response is a draft of the conversation, not just of the content. Each response shows you what Copilot understood from your prompt, so the gap between what you got and what you wanted tells you exactly which part of the prompt to fix. Treat prompting as a loop, not a one-shot request.
Microsoft's prompt guidance expects back-and-forth: if you're not satisfied, revise and try again. The skill the exam tests is diagnosing a weak response and choosing the right refinement, instead of starting over or blaming the tool.
Step 1: Check the prompt's structure. Map the response back to the four GCSE elements from 4.1. Only the Goal is required, but each missing element shows up in the response in a predictable way:
| What's wrong with the response | Likely missing element | Refinement |
|---|---|---|
| Generic, could apply to any company | Context | Add the situation, the purpose, and who it's for |
| Facts don't match your data | Source | Reference the file, email or meeting explicitly |
| Wrong length, format or tone | Expectations | State the format, length, tone and audience |
| Answers a different question | Goal | Restate the task precisely, as one clear instruction |
Step 2: Mind the wording. Two habits from Microsoft's guidance change results noticeably:
- Order matters. Copilot gives more weight to later parts of a prompt, so put the most important instruction last or restate it at the end.
- Say what to do, not what not to do. "Use plain language for a non-technical audience" works better than "Don't be technical."
Step 3: Evaluate the output before you refine. Microsoft suggests checking four qualities: clarity, accuracy, tone and audience fit, and context (does it cover what matters?). You can ask Copilot to help with this using a follow-up prompt, for example:
- "Highlight any claims in this section that are not supported by the information provided."
- "Does the tone of this draft match an executive audience?"
Treat that feedback as an additional perspective, not a final decision. You still own the judgment (Phase 3).
Step 4: Iterate with follow-ups. You don't have to rewrite the whole prompt. A short follow-up in the same chat ("Shorten this to five bullets and lead with the budget risk") keeps the context and fixes one thing at a time.
⚠️ Exam Trap: When a response is poor, the reflex answer "start a new chat and write a longer prompt" is usually wrong. The better answer names the specific missing element (context, source or expectations) or uses a targeted follow-up in the same chat. And asking Copilot to review its own draft is a useful refinement step, but it isn't a substitute for human verification of facts.
Reflection Question: Copilot's draft announcement is accurate but reads as stiff and overly technical for the all-staff audience it's meant for. Which element of the prompt was weak, and what follow-up prompt would fix it without losing the accurate content?
