Custom skills for Claude Code.
A structured debugging workflow that guides Claude through systematic issue investigation instead of jumping to fixes.
Flow: INTAKE → TRIAGE → REPRODUCE → HYPOTHESIZE → INVESTIGATE → FIX → VERIFY → CLOSE
- Bayesian-style hypothesis scoring (20/40/60/80) with evidence-based updates
- Severity triage (BLEEDING / DEGRADED / STABLE) to prioritize mitigation vs. investigation
- Loop-back protection — if a fix fails, it returns to investigation instead of thrashing
- TDD-aligned fix phase: write failing test, verify red, fix, verify green
- Resumable state via
/rca --resume
There are two ways to trigger the skill:
1. Slash command
Type /rca directly in your Claude Code prompt. You can optionally pass context or flags:
| Command | What it does |
|---|---|
/rca |
Scans the current conversation for errors and asks what you want to investigate |
/rca <description> |
Starts an investigation with the context you provide (error messages, stack traces, symptoms) |
/rca --resume |
Picks up a previous investigation from where it left off |
2. Natural language
You don't need the slash command — just describe the problem and Claude will activate the skill automatically. Examples:
- "My tests are passing locally but failing in CI"
- "Users are getting 502s on the /api/checkout endpoint"
- "The build broke after the last merge, can you figure out why?"
- "This function is returning null but I don't know why"
The skill triggers whenever Claude detects something is broken, failing, or behaving unexpectedly. The slash command gives you more control (e.g., --resume), but plain English works for starting a fresh investigation.
Once activated, Claude will:
- Triage the severity — is this actively hurting production, or a stable dev issue?
- Reproduce the error by observing actual output (not just reading code)
- Generate hypotheses with confidence scores (20/40/60/80)
- Investigate each hypothesis, updating scores as evidence confirms or eliminates them
- Fix the root cause with a failing test first, then a minimal fix
- Verify by running the full test suite — no "should work now" claims without proof
- Close with a commit linking the root cause and evidence chain
At each phase, Claude shows a hypothesis scoreboard so you can see what's being considered and why:
-- RCA Hypotheses ----------------------------------------
H1-DbPool ########.. 80 ^ <- connection timeout in logs
H2-AuthMW ##........ 20 v <- middleware unchanged since March
H3-RaceCondition ####...... 40 -> <- concurrent requests untested
----------------------------------------------------------
Copy any skill directory into your Claude Code skills folder:
# Clone the repo
git clone https://github.com/MP2EZ/claude-code-skills.git
# Copy the rca skill
cp -r claude-code-skills/rca ~/.claude/skills/rcaThen use /rca or describe a problem in any Claude Code conversation.
MIT