← ClaudeAtlas

diagnoselisted

Disciplined diagnosis loop for hard bugs and performance regressions. Reproduce → minimise → hypothesise → instrument → fix → regression-test. Use when user says "diagnose this" / "debug this", reports a bug, says something is broken/throwing/failing, or describes a performance regression.
JBSommeling/claude-config · ★ 0 · AI & Automation · score 70
Install: claude install-skill JBSommeling/claude-config
# Diagnose A discipline for hard bugs. Skip phases only when explicitly justified. When exploring the codebase, use the project's domain glossary to get a clear mental model of the relevant modules, and check ADRs in the area you're touching. ## Modes Two entry points set how far this skill runs: - **`/diagnose` — diagnose-only.** Run Phases 1–4 and 6; skip Phase 5. Confirm the root cause and report the diagnosis (root cause, the evidence that proves it, the hypothesis that held) — no fix, no regression test, no code changes. Phase 6 (Cleanup + post-mortem) runs automatically; the fix-verification items are N/A since no fix was applied. - **`/diagnose-fix` — diagnose and fix.** Run all six phases, including Phase 5. If the skill is triggered without either command, infer from the request: run Phase 5 only if the user clearly wants the bug fixed; otherwise stop after diagnosis. ## Phase 1 — Build a feedback loop **This is the skill.** Everything else is mechanical. If you have a fast, deterministic, agent-runnable pass/fail signal for the bug, you will find the cause — bisection, hypothesis-testing, and instrumentation all just consume that signal. If you don't have one, no amount of staring at code will save you. Spend disproportionate effort here. **Be aggressive. Be creative. Refuse to give up.** ### Ways to construct one — try them in roughly this order 1. **Failing test** at whatever seam reaches the bug — unit, integration, e2e. 2. **Curl / HTTP script** again