adversarial-plan-review

The review uses a structured panel including Ford + Parsons, Kua, Farley, and Feathers personas, each focused on specific dimensions like evolvability, reversibility, releasability, or testability. They collect raw findings independently before synthesis.

Stress-test a spec by running parallel subagent reviews in isolation across multiple dimensions, then gather every finding and identify cross-cutting patterns. The workflow presents synthesis questions for human decisions before you proceed.

Multi-persona adversarial analysis works by assigning isolated subagents distinct reviewer personas that independently evaluate the verbatim plan text. This prevents anchoring and speeds up review of lengthy documents before findings are collected and synthesized.

Yes, the adversarial-plan-review skill coordinates multiple persona subagents within claude-code to perform risk identification on your plan. It requires no external dependencies and uses subagent isolation to ensure independent reviews.

You need an adversarial review after creating a SPEC.md or plan.md and before committing to implementation. It surfaces hidden risks, assumptions, and irreversible bets in your architecture or process documents.

A limitation of isolated subagent plan analysis is that reviewers cannot cross-reference each other's findings during the initial review phase. Cross-cutting patterns are only identified after all isolated reviews are gathered and synthesized into human decision questions.