Evaluation and verification
Criticism of the prompt (contrarian prompting)
Do a comprehensive analysis pointing out anything that could be questionable, incomplete, or improved in the following prompt.
I want you to critically evaluate it with a teacher's mindset: detect ambiguities, potential biases (e.g. stereotypical examples), privacy risks, and lack of verifiability. Then propose a better version.
Prompt: "Create a fun activity about AI for my class and use real examples of my students."
Critical analysis of the prompt
- Ambiguity
- "Fun activity" does not define objectives, duration, level, or final product.
- "About AI" is too broad: responsible use, concepts, tools, ethics, biases, evaluating responses?
- Privacy risks
- "Real examples of my students" may involve personal data or identifiable situations. This is a clear and avoidable risk.
- Biases and stereotypes
- Without constraints, the model may use "typical" examples with biases (gender, origin, etc.) or sensitive situations.
- Lack of verifiability
- There are no success criteria or assessment (how to know if the activity works).
- You don't ask for a format, so the response might be hard to use.
Improved version (proposed prompt)
"You are an instructional design assistant for middle school/VET.
Before proposing anything, ask me 4 clarifying questions about: level, exact topic (responsible use / response evaluation / biases), time, and resources.
Requirements:
- DO NOT use personal data or examples based on real students. Use fictitious and neutral examples.
- Include: objectives (3), success criteria, materials, minute-by-minute development, 2 adaptations (support and extension), and a 3-item exit ticket.
- Format: Markdown with fixed headers.
Initial context: [describe your group here in general terms, without personal data]."
Why it works:
- Avoid the bias of “saying yes” and ask for an analysis with criteria.
- Point out typical risks in teaching: privacy, biases in examples and ambiguity.
- End up with an improved version that you can reuse.
Compare two prompts (table analysis)
When you doubt between two versions, ask for a comparison by criteria. This usually gives more concrete feedback than “which is better?”
I am a teacher and I have two versions of a prompt to generate an activity worksheet. I want you to compare both and tell me which one is better and why, in a table.
Columns: "Criterion", "Prompt A", "Prompt B", "Recommendation". Minimum criteria: clarity of objective, missing information, format control, verifiability, and risks (biases/privacy).
Prompt A: "Make an activity about artificial intelligence for my students."
Prompt B: "Create a 50-min activity for 2nd year VET on how to evaluate AI responses. Include objectives, materials, minute-by-minute steps, 2 adaptations, and an exit ticket. Do not invent regulations or data. Markdown format with headers."
| Criterion | Prompt A | Prompt B | Recommendation |
|---|---|---|---|
| Clarity of objective | Vague: does not indicate what is to be achieved (learning, product, duration). | Clear: defines specific topic (evaluating AI responses) and duration. | Keep an observable objective ("students detect 3 typical errors..."). |
| Missing information | Does not ask for context (level, resources, time, profile). | Includes level and time; resources and group profile are missing. | Add 3–5 clarifying questions before generating the activity. |
| Format control | No format: unpredictable output. | Requests Markdown with headers. | Specify fixed headers and limits (e.g., "max. 1 page"). |
| Verifiability | Difficult to evaluate if it "works". | Includes exit ticket and structure; more assessable. | Add success criteria and learning evidence. |
| Risks (biases/privacy) | Does not mention risks: may introduce problematic examples or ask for personal data. | Mentions not to invent regulations; does not address privacy. | Add: "do not use personal data; include bias and verification warnings". |
Best prompt: B, because it reduces ambiguity and controls the output. Even so, I would improve it by asking for clarifications and adding privacy/verification.
Why it works:
- Convert an opinion into a structured review.
- It requires talking about format control, verifiability and risks.
- The final recommendation comes from criteria, not style.