extended-rca

extended-rca

Stop writing shallow postmortems. A Claude Code plugin that runs a thorough, methodology-aware root cause analysis on resolved software incidents — and produces a structured engineering-grade postmortem that resists the usual RCA failure modes.

---

The problem

Most RCAs stop at the first plausible cause. The team finds a deploy that introduced a bug, writes a one-ticket postmortem, and moves on. Six weeks later, the same class of incident recurs — because "the deploy" was a trigger, not a root cause, and nobody looked at the contributing factors or the systemic conditions that made the trigger possible.

This plugin resists that pull. It gives Claude Code a disciplined workflow — 5-whys that don't stop at "human error," fishbone-style hypothesis generation across multiple categories, explicit layer separation (trigger / proximate / contributing / systemic), corrective actions tagged by type, and a self-check pass that catches hedged findings before handing the artifact over. The output is a postmortem your team can actually learn from.

Quick start

# 1. Clone into your Claude Code plugins directory
git clone https://github.com/sebdenes/ExtendedRCA ~/.claude/plugins/extended-rca

# 2. Restart Claude Code. Then in any project:
/rca we had a 45-minute checkout outage starting 14:02 UTC, 17% of requests failed...

That's it. Claude reads the skill, picks the right mode for your inputs, and produces a structured postmortem Markdown file.

What the skill actually does

Five modes, picked automatically or explicitly:

Mode	When to use	What you get
/rca <summary>	Default. You describe the incident or paste a timeline.	Full extended RCA with fishbone hypothesis generation, a 5-whys chain done properly, layered findings, and tagged actions.
/rca deepen <draft>	You already have a postmortem that stops too shallow.	Your draft + added alternative hypotheses, contributing factors, systemic causes, and tagged corrective actions.
/rca review <postmortem>	You just wrote one and want it graded, not rewritten.	A findings list scoring 10 axes out of 30, with specific issues cited and a recommended next step (share / deepen / rewrite).
/rca hypotheses <summary>	You want help thinking broadly before committing to a story.	At least three candidate hypotheses, fishbone-categorized, with confirm/disconfirm evidence notes.
/rca trends <folder>	You have N postmortems and want recurring systemic themes identified.	A synthesis document with 3–5 recurring themes, evidence counts, and recommended org-level interventions.

Why this resists real failure modes

Common RCA failure	How extended-rca resists it
Single-cause narrative ("the deploy broke it")	Forces multi-hypothesis generation across 6 fishbone categories (People / Process / Code / Infrastructure / Data / External)
Why-chain ends at "human error"	The skill and references/five-whys.md explicitly reject human error as a terminal cause
Layer confusion (trigger/proximate/systemic smooshed into "root cause")	The artifact template has separate sections for each layer
Action list is all "Prevent"	Actions are tagged Prevent / Detect / Mitigate / Respond; a healthy set spans ≥3 buckets
Hedged systemic findings ("communication could be improved")	Phase 6 self-check catches hedged phrasing and forces specific, non-blameful rewrites
Blame-ful phrasing toward a person or team	Self-check catches it before handover; the cross-team example shows the blameless-but-specific pattern
No record of rejected hypotheses	A dedicated "Hypotheses considered and ruled out" section
Postmortem reads as purely damning	A "What went well" section to preserve the patterns that worked
Single-incident tunnel vision	/rca trends surfaces recurring systemic themes across a folder of postmortems
Silent regression of skill quality	evals/ contains a structured suite of cases and a rubric for catching regressions

What you get in the output

A Markdown postmortem with deliberately-structured sections (see skills/extended-rca/references/artifact-template.md):