Purpose
Turn raw field failure and warranty return data into a ranked, evidence-based picture of what is actually failing and why — distinguishing systematic failure modes (design, supplier, process) from random incidents, and flagging vintage-correlated and environment-correlated patterns that point to a specific root cause. Solves the problem of warranty data that is high-volume but undifferentiated, where the real drivers are buried under noise.
When to Use
Invoke this skill when the user asks to:
- Investigate a sudden warranty claim spike or field return increase
- Find patterns in a batch of field failure, RMA, or warranty records
- Determine whether a failure is systematic or random
- Identify which parts or failure modes drive the majority of returns
- Check whether failures correlate with a production date range, lot, or build vintage
- Prepare a failure-mode summary for a corrective action or 8D investigation
Inputs Required
Before starting, collect:
- Product line/model: The product or family the failures belong to
- Failure records: Raw data — ideally including part number, failure mode/description, date of failure, date of manufacture or build vintage, and any environment/usage fields
- Population baseline: Units shipped or in service for the same period (needed for failure rate, not just counts)
- Time window: The date range of failures being analyzed
- Known failure modes: Any already-recognized modes or open investigations to reconcile against
- Available covariates: Which fields exist in the data — environment, region, usage hours/cycles, lot, supplier
- Target/threshold: Any warranty rate, PPM, or reliability target to measure against
If population baseline or manufacture dates are missing, analyze on raw counts and clearly flag which conclusions require normalized rates to confirm.
Instructions
Step 1: Data Inventory and Normalization
Establish what can and cannot be concluded from the data provided.
- List the fields present and their completeness (e.g., "failure mode populated on 82% of records")
- Note missing dimensions that limit analysis (no manufacture date → no vintage analysis; no population → counts only, not rates)
- Standardize failure mode language — collapse free-text variants into consistent mode categories (e.g., "won't power on," "dead," "no boot" → "no power-up")
- State the analysis basis explicitly: counts, or failures per units-at-risk (PPM / % of population)
Step 2: Pareto Analysis by Part and by Failure Mode
Identify the vital few drivers.
- Rank failures descending by part number — report count, share of total, and cumulative %
- Rank failures descending by failure mode — same three columns
- Identify the 80% cutoff line: which parts/modes account for the top 80% of failures
- Cross-tabulate the top parts against top modes to find the dominant part-mode pairs (this is usually where root cause lives)
Step 3: Vintage / Build-Date Correlation
Determine whether failures cluster in a production window.
- Bucket failures by manufacture date (monthly or by lot)
- Normalize by units built in each bucket — a raw count spike may just track higher production
- Flag any vintage with failure rate materially above the baseline (call out >2x baseline as a strong lot/change signal)
- Correlate elevated vintages against known engineering changes, supplier changes, or process events if provided
Step 4: Environment and Usage-Profile Clustering
Test whether failures concentrate under specific conditions.
- Segment by available covariates: region/climate, usage intensity (hours, cycles, duty), installation type, customer segment
- Compare the failure population's distribution against the installed-base distribution — over-representation is the signal (e.g., "38% of failures from humid-climate units, but only 14% of the fleet ships there")
- Identify usage-profile thresholds where failure rate accelerates (early-life, wear-out, or duty-driven)
Step 5: Systematic vs. Random Classification
Apply a clear decision logic to each significant mode.
Classify a failure mode as systematic when two or more of these hold:
- Concentrated on a specific part-mode pair (Pareto driver)
- Correlated with a build vintage, lot, or supplier
- Correlated with an environment or usage condition beyond installed-base proportion
- Rate is trending up over time or exceeds target/threshold
Classify as random / background when:
- Distributed evenly across parts, vintages, and conditions
- Low rate near or below expected background reliability
- No time trend
For time trends, characterize the failure pattern shape: infant mortality (early-life, declining hazard), constant/random (flat hazard), or wear-out (rising hazard with age/usage).
Step 6: Ranked Findings and Recommended Next Actions
Produce prioritized conclusions.
- List each systematic pattern found, ranked by contribution to total failures, with the evidence that classifies it as systematic
- For each, state the most likely root-cause direction (design / supplier / process / usage) based on which correlations fired
- Recommend the specific next diagnostic step per pattern (e.g., "pull 5 field returns from the July–Sept vintage for teardown," "request supplier lot traceability for connector P/N")
- Separate out the random-background residual and note it does not warrant corrective action
Output Format
- Structure: Diagnostic report in the six-step order, with Pareto results as ranked tables and findings as a prioritized list
- Length: 900–1,300 words
- Format: Structured document with tables for Pareto and vintage analysis; bulleted findings; prose only for interpretation
- Sections: Data Inventory & Basis → Pareto by Part & Mode → Vintage Correlation → Environment/Usage Clustering → Systematic vs. Random Classification → Ranked Findings & Next Actions
- Tone: Quality engineer — evidence-first, cautious about overclaiming from thin data, explicit about confidence and limitations
Quality Checklist
Before delivering output, verify:
- [ ] Analysis basis (counts vs. normalized rate) is stated, and any count-only conclusions are flagged as requiring population data to confirm
- [ ] Pareto tables include count, % of total, and cumulative % — with the 80% cutoff identified
- [ ] Vintage analysis normalizes against units built, not raw counts alone
- [ ] Environment/usage clustering compares failure distribution against installed-base distribution, not just failure counts
- [ ] Every mode is explicitly classified systematic or random using the stated decision logic
- [ ] Each systematic pattern has a specific, actionable next diagnostic step — not "investigate further"
Notes
- The single most common analytical error here is mistaking a count spike for a rate spike — a lot that shipped 10x volume will show 10x failures at identical quality. Always normalize before concluding.
- The dominant part-mode pair from the cross-tab (Step 2) is where teardown effort should go first; it is almost always more productive than chasing the long tail.
- Free-text failure descriptions must be normalized before Pareto ranking, or the same mode will fragment across categories and hide the true top driver.
- If manufacture dates exist but no engineering change log is provided, still flag the elevated vintage window precisely — that date range is what the user takes back to their change-control records.
How to use this skill
This file is a Claude Code skill in Anthropic's standard format — markdown with YAML frontmatter at the top.
Installing in Claude Code:
~/.claude/skills/<folder-name>/SKILL.md— auto-loads in every session anywhere on your machine (recommended for skills you'll use across multiple projects)<your-project>/.claude/skills/<folder-name>/SKILL.md— auto-loads only when working in that project (recommended for project-specific skills)
Restart Claude Code or open a new session to register.
Using in Claude.ai, ChatGPT, Gemini, or other AI tools:
- Drag this file into a new chat, or copy and paste its contents
- Send your task or request as your next message — the AI will apply the skill where it fits
For Claude Code skill details: https://docs.anthropic.com/en/docs/claude-code/skills