phase-gated-debugging
Utilisez lors du débogage de tout bug. Applique un protocole en 5 phases où les modifications de code sont bloquées jusqu'à la confirmation de la cause racine. Prévient les tentatives de correction prématurées.
Le contenu de ce skill est dans sa langue d’origine (souvent l’anglais).
Phase-Gated Debugging
Overview
AI coding agents see an error and immediately edit code. They guess at fixes, get it wrong, and spiral. This skill enforces a strict 5-phase protocol where you CANNOT edit source code until the root cause is identified and confirmed.
Based on claude-debug (full plugin with PreToolUse hook enforcement).
When to Use
Use this skill when:
- a bug keeps getting "fixed" without resolving the underlying issue
- you need to slow an agent down and force disciplined debugging before code edits
- the failure is intermittent, a regression, performance-related, or otherwise hard to isolate
- you want an explicit user confirmation checkpoint before any fix is applied
The Protocol
Phase 1: REPRODUCE
Run the failing command/test. Capture the exact error. Run 2-3 times for consistency.
- Do NOT read source code
- Do NOT hypothesize
- Do NOT edit any files
Phase 2: ISOLATE
Read code. Add diagnostic logging marked // DEBUG. Re-run with diagnostics. Binary search to narrow down.
- Only
// DEBUGmarked logging is allowed - Do NOT fix the bug even if you see it
Phase 3: ROOT CAUSE
Analyze WHY at the isolated location. Use "5 Whys" technique. Remove debug logging.
State: "This is my root cause analysis: [explanation]. Do you agree, or should I investigate further?"
WAIT for user confirmation. Do NOT proceed without it.
Phase 4: FIX
Remove all // DEBUG lines. Apply minimal change addressing confirmed root cause.
- Only edit files related to root cause
- Do NOT refactor unrelated code
Phase 5: VERIFY
Run original failing test — must pass. Run related tests. For intermittent bugs, run 5+ times. If verification fails: root cause was wrong, go back to Phase 2.
Bug-Type Strategies
| Type | Technique |
|---|---|
| Crash/Panic | Stack trace backward — trace the bad value to its source |
| Wrong Output | Binary search — log midpoint, halve search space each iteration |
| Intermittent | Compare passing vs failing run logs — find ordering divergence |
| Regression | git bisect — find the offending commit |
| Performance | Timing at stage boundaries — find the bottleneck |
Key Rules
- NEVER edit source code in phases 1-3 (except
// DEBUGin phase 2) - NEVER proceed past phase 3 without user confirmation
- ALWAYS reproduce before investigating
- ALWAYS verify after fixing
Limitations
- Use this skill only when the task clearly matches the scope described above.
- Do not treat the output as a substitute for environment-specific validation, testing, or expert review.
- Stop and ask for clarification if required inputs, permissions, safety boundaries, or success criteria are missing.