Run a Hostile Second Pass on an Answer Before You Rely on It

You have an AI answer you are about to act on, and a faint suspicion it is too tidy. This runs a hostile second pass: the model switches sides, attacks every claim and assumption in the answer, then tells you what survived. Cheaper than finding out in the meeting.

<context>
You are about to switch roles. {ANSWER_TO_TEST} is an AI-generated answer the user intends to rely on for [what you will use it for]. Your job is to attack it as a hostile, well-informed reviewer, then give a fair verdict.
</context>

<task>
**Pass 1, understand:**
1. Restate what the answer claims and what it recommends, in neutral terms.

**Pass 2, attack:**
2. As a sceptical domain expert, challenge every significant claim: what evidence is missing, what is asserted rather than shown, what a competitor or auditor would seize on.
3. List the assumptions the answer silently relies on, and for each say what happens if it turns out to be false.
4. Name at least two realistic scenarios where following this answer goes wrong.

**Pass 3, verdict:**
5. Separate the findings: what survived the attack, what needs verification before use, what should be removed or rewritten.
6. Produce a revised version of the answer that would survive the same attack, with the fixes applied. If the original is beyond saving, say so and explain what to ask for instead.
</task>

<output_format>
- Restatement: one short paragraph
- Attack: bullets grouped under claims, assumptions, failure scenarios
- Verdict: three lists (survived, verify first, remove or rewrite)
- Revised answer: full text, ready to use
- Tone: blunt in the attack, fair in the verdict
</output_format>