← ClaudeAtlas

bug-investigationlisted

Use when investigating a reported bug, diagnosing a failure, or doing root-cause analysis — especially before concluding "not a bug", tracing a symptom, or explaining unexpected behavior. Preloaded on investigation agents.
lklimek/claudius · ★ 1 · AI & Automation · score 67
Install: claude install-skill lklimek/claudius
# Bug Investigation Default discipline for diagnosing a reported bug. The user's reproducible observation is ground truth — the analysis must explain it, not explain it away. ## Rules 1. **Observation over theory** — the user's reproducible facts are ground truth to explain; their causal explanation is only a hypothesis. Refuting the theory ≠ explaining the observation. (User imprecision usually lives in the explanation, not the observation.) 2. **Entry point over name** — statically trace the call graph from the actual UI/CLI entry point (the thing the user clicked/ran) down to the outcome. Never anchor on a function whose name merely matches the feature. 3. **The exercised path, not a correct path** — when ≥2 plausible code paths exist, verify the one actually hit. Proving a correct path exists is not proving the user's path is correct. 4. **Reproduce or it's unsolved** — analysis that cannot reproduce the user's concrete observation is INCOMPLETE. Never conclude "not a bug" until the observation is explained. A clash between analysis and a user-observed fact is a STOP signal, not a footnote. When the repro artifact is a test, it must be RED against the documented contract on the buggy revision, not green mirroring the defect — see `coding-best-practices` § Workflow Discipline TDD "Repro tests go RED first". 5. **Brief with the literal reproduction** (coordinator-facing) — investigation spawn prompts MUST quote the user's exact reproduction steps and the literal entry po