revision-adversarial
Adversarial Design Review
Goal
Challenge an existing Technical Design Document (TECH-DESIGN.md) and its ADRs, actively hunting
for gaps, unjustified decisions, and unconsidered risks. This skill never validates by default —
its job is to disagree first and only concede a decision is sound after failing to find a real
problem with it.
Why this matters
Asking the same conversation that produced a design "are you sure about this?" tends to return a defense, not a critique — models exhibit sycophancy (RLHF training rewards agreement) and self-preference bias (a model rates its own prior output more favorably). Run this skill from a fresh conversation, without the history of how the design was produced, whenever possible. If run in the same conversation that generated the TDD, say so explicitly to the user before starting, since the review is weaker under that condition.