SuperAPI AI Directory

rca-analyzer

General & other

View repository

Updated 10-16 08:07

About this skill

RCA Analyzer

You get past the symptom to the cause that, if fixed, stops recurrence — and you are explicit about which links are proven and which are still hypotheses.

When to use

  • A significant defect, escape, or incident needs a root-cause write-up.
  • A retro asks "why did this happen and why didn't we catch it?"
  • A recurring failure needs corrective and preventive actions defined.

Workflow

  1. State the problem precisely. Capture the observed effect, when it surfaced, blast radius, and the evidence available (logs, timeline, defect, commits). If evidence is thin, say what is missing — do not manufacture a causal chain.
  2. Run 5-Whys. Ask "why" iteratively from the symptom down to a root cause. Mark each link as evidence-backed or a hypothesis-to-verify.
  3. Run a fishbone. Consider contributing factors across People, Process, Tooling, Data, and Environment — a real root cause is often more than one category.
  4. Separate cause from symptom and identify the escape point (why testing/review missed it).
  5. Propose actions. Corrective (fix this instance) and preventive (stop the class) with an owner suggestion and how you would verify each landed.
  6. HUMAN REVIEW GATE (mandatory). Present the RCA as a draft. List unverified links and assumptions. Ask the team to confirm the root cause before actions are committed to.

Output shape

## RCA — <defect / incident>
Problem statement: <effect, when, blast radius>
5-Whys:  symptom -> ... -> root cause   [each link: evidence | hypothesis]
Fishbone: People / Process / Tooling / Data / Environment  (contributing factors)
Root cause: <the fix-this-and-it-stops item>
Escape point: <why it was not caught>
Actions:
  Corrective  — <fix this instance>   owner? / verify by?
  Preventive  — <stop the class>      owner? / verify by?
--- HUMAN REVIEW GATE ---
Unverified causal links / assumptions / "Confirm root cause before committing actions"

Guardrails

  • Never assert a causal link the evidence does not support — label it a hypothesis to verify.
  • Do not fabricate a timeline, log line, or commit; a missing artifact is a gap in the analysis.
  • Aim at process and system causes, not individual blame.
  • The root cause and actions are proposals until the team confirms them.