Rca
Root cause analysis for a bug or incident. Fans out debuggers and researchers in parallel to find what's actually causing a problem, converges on a verified root cause with evidence, and sketches a remediation plan — without implementing it. Ends by recommending /drive if there's a reasonable fix approach. Use when something is broken and you need to understand WHY before touching code.From its SKILL.md
npx -y skills add cazzer/cc-team-skills --skill rcaAssembled from the repository path, not quoted from the project. Check it against their README if it does not work.
2 things to look at
- no licenseNo license file was found in the repository. Code published without one is not open source by default, so using it at work is a question for whoever answers licensing questions where you are.
- 1 stars1 stars. Stars are a popularity signal and not a quality one, but at this level it is likely that nobody has read this closely except its author, and you would be relying on your own review.
SKILL.md
6.9 KB, as published. Nobody here has run it
RCA — Root Cause Analysis
You are running a root cause investigation. Your job is to understand why a problem is happening — deeply, with evidence — and to hand off a remediation plan. You do not fix it. No edits, no implementation. Investigation and planning only.
Preflight
!gh auth status >/dev/null 2>&1 && echo "MODE: gh CLI (authenticated) — use gh for all issue/PR ops." || echo "MODE: GitHub MCP — gh is unavailable or unauthenticated. This is NORMAL in a cloud session and is NOT a failure: use the GitHub MCP tools for every issue/PR read and write, and treat any gh command in this skill as a spec of intent, not a literal command. Find the tools with ToolSearch (query: \"github issue pr comment\") to load their schemas before calling. If NEITHER gh nor GitHub MCP is available, report that plainly — never silently skip an issue/PR step."
Project context
!cat CLAUDE.md 2>/dev/null || echo "No project context found — proceeding without project-specific context"
Researcher role (inject into research subagents)
!cat .claude/skills/context/roles/researcher.md 2>/dev/null || cat "${CLAUDE_PLUGIN_ROOT}/context/roles/researcher.md" 2>/dev/null || cat "$HOME/.claude/skills/context/roles/researcher.md" 2>/dev/null || echo "WARNING: Researcher role not found"
Skill handoffs
!cat .claude/skills/context/handoffs.md 2>/dev/null || cat "${CLAUDE_PLUGIN_ROOT}/context/handoffs.md" 2>/dev/null || cat "$HOME/.claude/skills/context/handoffs.md" 2>/dev/null || echo "WARNING: Handoff map not found at expected path"
Subagent mapping (resolve investigator slots)
!cat .claude/routing.md 2>/dev/null || echo "No routing table — baked-in role profiles in context/roles/ are the default; investigation works with local roles + Explore. Run /setup-agents only to add optional plugin agents."
How RCA works
Parallel investigation → converge on a verified cause → sketch remediation → hand off. Read-only throughout.
Step 1 — Frame the problem
Capture the symptom precisely before theorizing:
- If given a
#issue, load it (gh issue view). If given error text or a stack trace, quote it exactly. - Nail down: what's the observed behavior vs expected, when did it start, what's the blast radius (one user / one surface / systemic), is it reproducible, what changed recently (
git log, recent PRs). - Restate the problem in one or two lines so the investigators share a target.
If the symptom is too vague to investigate (no repro, no error, no signal), that itself is the finding — say what's needed to make it investigable, and stop. Don't fan out on a guess.
Step 2 — Form distinct hypotheses
List 2–4 candidate causes, each a different angle — not restatements of the same guess. Each becomes an investigator. Typical angles:
- Code path — trace the logic; where could it produce this behavior? (debugger)
- Regression — did a recent change introduce it?
git log/blame/bisect the suspect range. (debugger) - Data / logs / traces — what do error signatures, logs, or runtime data say? (error-detective / observability)
- Environment / contract — schema, RLS, env vars, API contract, dependency version drift. (researcher / db / backend)
Step 3 — Fan out investigators (parallel, read-only)
Spawn one subagent per hypothesis in a single message so they run concurrently. Resolve each slot via the routing table above (debugger → debugging-toolkit:debugger, researcher → Explore, error/log angle → incident-response:error-detective, data-layer → db-specialist).
Every investigator prompt MUST include:
- The framed problem (Step 1) + the specific hypothesis this agent owns.
- The relevant
docs/engsection(s),features/<feature>/folder(s), and files in play. - A hard constraint: read-only. Investigate and report evidence for/against this hypothesis with file:line citations and a confidence level. Do NOT edit any files, do NOT implement a fix.
Ask each to return: verdict (supported / refuted / inconclusive), the evidence, and any adjacent cause it stumbled on.
Step 4 — Converge on the root cause
Synthesize the returns. Then try to disprove your leading candidate — a plausible-but-wrong cause that survives to remediation is the failure mode here. Distinguish:
- Root cause — the actual origin, with evidence (file:line, commit, log).
- Contributing factors — what made it worse or masked it.
- Confidence — high / medium / low, and what would raise it.
If investigators conflict or all come back inconclusive, spawn a targeted second round on the gap rather than guessing.
Step 5 — Sketch the remediation plan (do not implement)
For the confirmed cause, outline the fix — approach only, no code:
- The fix approach (and 1–2 alternatives if there's a real tradeoff), with a recommendation.
- Scope / blast radius: which surfaces and files, how big.
- Risk and what could regress; what tests or verification the fix needs.
- Any prerequisite (data migration, backfill, config) before the code change.
Step 6 — Hand off
- If there's a reasonable approach (confidence not low, scope understood): recommend launching
/drivewith a one-line remediation framing, e.g. "Recommended next step:/drive fix <root cause> in <feature> — <approach>." Then stop and wait for the user to approve. Do NOT invoke/driveyourself — surface it as the suggested next step and let the user pull the trigger. Drive will size and route it (tweak/focus/sprint) once they do. - If confidence is low or the fix is unclear: say so plainly. Recommend what's needed first — more repro, more logging, a spike — not a fix. Do not push to
/driveon a guess.
Never implement the fix from within RCA, and never auto-launch /drive. The handoff — a verified cause plus a recommended path — is the deliverable; running it is the user's call.
Anti-patterns
- Don't edit code or implement the fix. RCA stops at the plan.
- Don't stop at the first plausible cause. Verify — try to refute it before you trust it.
- Don't fan out investigators on the same guess worn three ways. Distinct hypotheses only.
- Don't let investigator subagents modify files — they're read-only; state that constraint in every prompt.
- Don't hand off to
/driveon a low-confidence guess. A wrong root cause sends drive down the wrong path. - Don't theorize past a missing repro. No signal → say what's needed to get one, and stop.