gabrielmoreira RMA Failure Analysis Skill
Raw RMA data lies: it mixes buyer's remorse, user error, and shipping damage in with real design and manufacturing defects. This skill turns returns into decisions — triage first so the Pareto is of verified failure modes, run the top modes through 8D discipline (contain now, root-cause properly, correct permanently), and price each mode in cost-of-quality terms so the fix competes for resources on money, not anecdote.
What This Skill Produces
- A triaged breakdown of returns using a standard RMA taxonomy
- A Pareto of verified failure modes with rates against units shipped
- An 8D-structured analysis for each top failure mode
- Cost-of-quality framing: cost per return, per mode, and the fix ROI
- Prioritised corrective actions with owners and cut-in points
Required Inputs
Ask for these if not provided; analyse whatever slice exists, but state the denominator caveats plainly:
- RMA records — return reasons, dates, symptoms, any teardown/FA findings
- Units shipped per period — the denominator; return counts without it are useless
- Product age mix — manufacture date or batch, to separate infant mortality from wear-out
- Cost inputs — per-return logistics, refurb/scrap cost, support cost per case (estimate and label if unknown)
- Known changes — ECOs, factory or component changes that bracket the data in time
Analysis Framework
Step 1 — Triage taxonomy. Bucket every return before any Pareto:
| Bucket | Meaning |
|---|---|
| NTF / CND | No trouble found — unit passes full test; count separately, it's a UX/expectation signal |
| CID | Customer-induced damage (drop, liquid) — a robustness signal, not a defect |
| OBF / DOA | Failed out of box — points at outgoing quality or transit |
| SW-resolvable | Fixed by update/reset — cheapest class to kill |
| Verified HW failure | Real defect, classified by subsystem and failure mode |
| Remorse / non-technical | Returned working — exclude from quality analysis, report separately |
Step 2 — Pareto verified failures only, by failure mode (not symptom — "won't charge" is a symptom; "USB connector solder crack" is a mode). Express each as % of units shipped in the exposed population, with the time window stated.
Step 3 — 8D per top mode (top 3–5 carry most of the cost): D1 team · D2 problem statement with data · D3 containment (screen stock, hold lots, factory rescreen — dated) · D4 root cause via evidence (teardown, cross-section, batch correlation), labelled [verified] or [hypothesis] · D5 corrective action chosen · D6 implementation with cut-in (ECO/date/serial break) · D7 recurrence prevention (test coverage, DFM rule, spec change) · D8 closure criteria (return rate for the mode falls to X by date Y).
Step 4 — Cost of quality. Cost per return = freight + refurb/scrap + support labour + replacement unit margin. Annualise per mode; compare fix cost vs failure cost; note warranty-accrual impact.
Output Format
RMA failure analysis: [product] — [period]
- Summary — return rate vs target, headline modes, the one-paragraph verdict
- Triage breakdown — table: bucket, count, % of returns, % of shipped
- Verified-failure Pareto — mode, count, rate vs shipped, trend, batch correlation
- 8D per top mode — the eight disciplines, with D4 evidence labelled verified/hypothesis
- Cost of quality — per-return cost, annualised per mode, fix ROI table
- Actions — containment (now) and corrective (cut-in) with owners and dates
- Data caveats — denominator gaps, lag effects, unteardown returns
Quality Checks
- Every rate has a denominator and an exposure window — counts alone never appear
- NTF, CID, and remorse are separated out before the failure Pareto
- Pareto items are failure modes, not symptoms
- Every root cause is labelled
[verified]or[hypothesis]with its evidence - Containment actions are distinct from corrective actions, each dated and owned
- Cost of quality uses stated inputs; estimates are labelled as estimates
Anti-Patterns
- Do not Pareto raw return reasons — triage first, or NTF and remorse will drown the real defects
- Do not report return counts without units shipped and the exposure window
- Do not close an 8D at D5 — a corrective action without cut-in verification and recurrence prevention is a wish
- Do not treat NTF as noise to discard — a high NTF rate is a product or support failure of its own
- Do not root-cause by vote — teardown evidence and batch correlation, or label it a hypothesis
- Do not compare return rates across cohorts with different time-in-field — young cohorts always look better