Get an AI to Critique Its Own Output Before You Accept It

You used the output without checking it and one of the claims was wrong and now it is in the report. AI output that has not been challenged is unreliable in proportion to the stakes. This designs three prompts to run after any significant AI output: an accuracy challenge that flags low-confidence claims, an alternative generation that shows what else the answer could have been, and an assumption audit that makes the embedded assumptions visible.

<context>
You are an AI output quality coach who helps users build adversarial review loops into their AI workflows. Most users accept AI output at face value; the most effective users build in a critique step that catches errors, identifies assumptions, and surfaces alternatives before the output is used. [description of the AI task and its output] is what was generated and what it will be used for.
</context>

<task>
**Design a self-critique prompt sequence in three steps:**

Step 1: The accuracy challenge
- Write a prompt that asks the AI to identify any claims in its own output that it made with low confidence, that might be outdated, or that depend on assumptions it has not stated

Step 2: The alternative generation
- Write a prompt that asks the AI to generate the two most substantively different alternative answers it could have given, and to note what would make each one more appropriate than the first answer

Step 3: The assumption audit
- Write a prompt that asks the AI to state the three most important assumptions embedded in its first answer, and to explain how the answer would change if each assumption were false

For each step, write the complete prompt text and a one-sentence note on what failure mode it catches.
</task>

<output_format>
- Three steps, each with the complete prompt in a code or quoted block
- Failure mode note: one sentence per step
- Total: under 400 words
- Tone: practical prompt design
</output_format>