What is an adversarial review?

An adversarial review is a second AI pass whose only job is to attack the first output. Instead of asking Claude to review something, which tends to produce praise, you instruct it to act as a skeptic that assumes the work is wrong and has to find the flaws.

What is the Claude adversarial review prompt?

A copy-paste instruction that tells Claude to switch sides and critique an output as a hostile reviewer: no praise, a severity label and a concrete fix for every finding, and a clear verdict at the end. The full prompt is on the page, free to copy.

How is adversarial code review different?

Same idea aimed at a diff. The code-review variant attacks correctness, security, concurrency, performance, and test coverage, and ends with APPROVE, REQUEST_CHANGES, or BLOCK. It is a fast second opinion before you merge, not a replacement for human review.

Does it work in ChatGPT and other models?

Yes. The prompt is model-agnostic, so it works in ChatGPT, Claude, Gemini, or Copilot. For the sharpest result, run the review in a fresh chat or a different model so the reviewer does not inherit the first pass assumptions.

Adversarial Review Prompt: Make Claude Attack Its Own Answer

Three copy-paste prompts that turn Claude into a hostile reviewer of its own output. One for writing, one for code, one as a Claude Code slash command.