What problem does it solve? Confident answers are not always correct ones, and long sessions quietly turn assumptions into facts. This Skill catches wrong directions in non-trivial decisions — branching logic, cross-boundary changes, unverifiable invariants, irreversible operations — while course-correction is still cheap, instead of discovering them in production. ## Core Features & Use Cases - Five-step doubt cycle: CLAIM the decision, EXTRACT the smallest reviewable artifact plus its contract, DOUBT it with a fresh-context adversarial reviewer, RECONCILE findings by precedence (contract misread, actionable, trade-off, noise), and STOP at a bounded condition (trivial findings, 3 cycles, or user override). - Adversarial reviewer prompting: Passes only the artifact and contract — never your conclusion — with an issues-only prompt so the reviewer disproves rather than validates. - Cross-model escalation: In interactive sessions it always offers a second opinion via external CLIs (Gemini, Codex) with read-only sandboxing, stdin piping to avoid shell injection, and explicit per-invocation user authorization. - Use Case: Before committing a new caching layer claimed to be thread-safe, extract the diff and its concurrency contract, spawn a fresh-context reviewer to hunt for races and unstated assumptions, then classify and act on the findings before merging. ## Quick Start Ask the agent to apply doubt-driven development to review the thread-safety claim of the new caching layer before committing it.
