Skip to content
SHRINESHRINE

Adversarial Review

A reviewer whose job is to find holes, not to approve.

The Pattern

Section titled “The Pattern”

After generating output, pass it to a separate agent with explicit instructions to:

  • Find flaws, gaps, or unstated assumptions
  • Challenge claims that lack evidence
  • Identify edge cases the original missed

The adversarial reviewer is not trying to be helpful to the original output. It’s trying to break it.

When to Use

Section titled “When to Use”
  • High-stakes decisions where being wrong is expensive
  • Output that will be trusted without human review
  • Claims that sound plausible but haven’t been verified
  • Any “are we sure?” moment

Implementation Notes

Section titled “Implementation Notes”
  • The reviewer should NOT see the original prompt or reasoning, only the output
  • Explicitly instruct it to default to skepticism
  • A finding that survives adversarial review is stronger than one that was never challenged

Anti-patterns

Section titled “Anti-patterns”
  • Using the same model instance for both generation and review (context bleeds)
  • Instructing the reviewer to “check if this looks good” (too soft)
  • Ignoring findings because they’re inconvenient
Section titled “Related Patterns”