← ClaudeAtlas

doubt-driven-reviewlisted

Adversarial fresh-context review for non-trivial decisions before they stand. Use for production-impacting logic, security-sensitive changes, unfamiliar code, or high-blast-radius architecture choices.
charlieviettq/awesome-agent-skill · ★ 14 · AI & Automation · score 81
Install: claude install-skill charlieviettq/awesome-agent-skill
# Doubt-driven review ## Non-trivial when - Branching logic or cross-module boundary changes - Properties types/compiler cannot verify - Irreversible deploy, migration, or public API change ## When not to use - Renames, formatting, obvious one-liners ## Cycle 1. **CLAIM** — decision + why it matters (2-3 lines) 2. **EXTRACT** — artifact + contract only (no reasoning journey) 3. **DOUBT** — fresh review with adversarial prompt ("find issues, do not validate") 4. **RECONCILE** — classify: contract misread / actionable / trade-off / noise 5. **STOP** — trivial findings only, 3 cycles max, or user says ship ## Reviewer input Pass **ARTIFACT + CONTRACT** only—not your CLAIM (avoids agreement bias). ## Reconcile precedence 1. Contract misread — fix contract, re-run 2. Valid actionable — change artifact 3. Valid trade-off — document explicitly 4. Noise — note and move on ## Optional second opinion Offer cross-model or `codex review` for high stakes; user decides. Never silent skip. ## Related `verify-before-done`, `test-first-development` (RED step as behavioral doubt), `gstack/review` (post-hoc PR)