← Back to library

Rank likely root causes and draft the 8D from plant data

Anthropic's Claude for Manufacturing plugin, published November 12, 2026, includes a skill for investigating a nonconformance after the NCR is written. It refuses to correlate data whose clocks it cannot line up, computes every figure instead of estimating it and returns ranked candidates rather than one verdict. We explain how to use it.

  • OutcomeUse the root-cause-capa skill from Anthropic's Claude for Manufacturing plugin to line up sensor, inspection, maintenance and lot records on one checked timeline, run control-chart and clustering checks in code, rank several candidate cause chains with evidence for and against each step, name the cheapest test that separates them and draft an 8D that leaves verification and closure for people to fill in.
  • SetupAn afternoon+
  • LevelAdvanced
  • Creditanthropics
View original
Works with
ClaudeClaude CoworkClaude Code

Published Nov 13, 2026Updated Nov 13, 2026AI dry run on sample input: passed Nov 13, 2026Source verified Nov 13, 2026

When should you use it?

  • A customer asks for a corrective action report after a quality escape
  • A defect keeps returning and the earlier fix clearly did not hold
  • Evidence sits in separate systems, such as a historian, the inspection database, the maintenance log and the ERP

Not for: Signing off verification or closure; the skill deliberately leaves those sections empty for a person; Issues with no measured characteristic, where a control chart cannot be computed; Naming an individual operator as the cause; the skill reports patterns, not people

What do you need, and what do you get?

You need

  • An approved NCR that cites the requirement that was missed
  • Exports covering the event window: inspection measurements, sensor history if any, maintenance records and lot or receipt data
  • Your plant profile with clock settings for each system and, ideally, your controlled 8D template

You get

  • Ranked candidate cause chains, each step with its supporting and contradicting evidence and a source
  • A confirm test and a kill test per candidate, plus the single cheapest test that separates them
  • A draft 8D with containment, root cause and corrective actions filled in and verification and closure left blank
  • An evidence appendix with the aligned timeline, computed control limits and the checks that could not run
  • When requested, a redacted customer version with a reviewable list of every removal

How does it work, step by step?

  1. Install and onboardAdd Claude for Manufacturing from the Claude plugin directory and complete onboarding so the plant profile records each data source's clock and your 8D template.
  2. Start from the NCRHand it the approved nonconformance report and the data exports for the period around the event, then ask for a root cause investigation.
  3. Check the timeline firstRead how it aligned the clocks. If it could not verify alignment, accept that time-based correlations were refused and review what still stands.
  4. Run the cheapest testUse the named test to separate the leading candidates before anyone commits to a corrective action.
  5. Finish the 8D yourselfReview the draft, approve any supplier-caused finding and any redactions, and complete verification and closure once the fix has been proven.

The prompt

Our adapted version, written to work in Claude, Claude Cowork and Claude Code. Values in lime are placeholders.

Fill in the blanks (optional; the copied prompt uses your values)

Prompt2 variables
Investigate the root cause of NCR {ncr_number}. The evidence is in {exports}.
First put every source on one timeline and tell me whether you could confirm the clocks line up. If not, do not draw any time-based correlations and say which conclusions still hold.
Compute control limits, run rules and group comparisons in code and show the figures.
Give me ranked candidate cause chains with evidence for and against each link, a confirm test and a kill test for each, and the cheapest test that separates the top candidates.
Draft the 8D with verification and closure left blank, and do not name an individual as the cause.

The original skill is published by anthropics on GitHub (anthropics/claude-plugins-official). Open it there to install or import it; the prompt above is our own adaptation.

What can go wrong?

  • Mixing local time and UTC across systems produces convincing but false correlations. Fill in the clock settings during onboarding.
  • Small samples look like patterns. Two failures on one shift and one on another is not evidence about that shift.
  • Blaming a supplier on a tie. A supplier-caused finding is a commercial claim, so settle the tie with a test before raising it.
  • The plugin is published as a reference implementation that Anthropic does not actively maintain, so try it on a closed case first.

Our take

Root cause work fails in two familiar ways: the team settles on the first plausible explanation, or someone reads a trend into a chart that is not really there. This skill is built to resist both. Every limit and ratio is calculated in code, every step in a proposed cause chain carries evidence for and against, and the output is a ranking rather than a verdict. The feature we like most is the refusal to correlate data when it cannot prove the timestamps line up. Mixed time zones between a sensor historian and an inspection system are a common source of false conclusions, and saying so openly is unusual. We also like that the verification and closure sections of the 8D stay empty, because filling them in before a fix is proven would be a false record. Cautions: it is only as good as the exports you give it, the plugin is labeled a reference implementation, and supplier blame should wait until a tie is broken by a test. In our dry run on a fictional bore-diameter escape, it ranked tool wear and a material lot change as a tie and proposed a hardness check to separate them.

Where the time goes back: Gathering data from four systems onto one timeline and running the statistics by hand can take a quality engineer days. The skill produces a ranked, sourced starting point and the next test to run, so the team spends its time confirming causes instead of assembling spreadsheets.

FAQ

Who made it and what license does it use?
Anthropic published it in the anthropics/claude-plugins-official repository on GitHub under the Apache 2.0 license.
Does it need live connections to our systems?
No. It can work from exported files alone. Connections to Microsoft 365 or an ERP such as Oracle NetSuite or Infor M3 simply add more evidence.
Will it close the corrective action?
No. It drafts the 8D, but the verification and closure sections stay empty until a person records a real result and signs.

Source and credit

This playbook is based on work by anthropics. We summarize it and link to it and write our own adapted prompt.

Original
root-cause-capa skill (anthropics/claude-plugins-official, Claude for Manufacturing)
Creator
anthropics
Published on
GitHub (anthropics/claude-plugins-official)
Original date
Nov 12, 2026
License
Apache-2.0
Originally for
Claude
AI dry run
Nov 13, 2026: passed. Dry-run tested by our AI test agent on sample input. Alignment verified first, figures computed, tie stated with the cheapest test, supplier flag held, verification and closure left blank.
AI dry run on sample data; not run with live connected accounts. The playbooks.bot test agent filled the prompt with synthetic inputs, produced the output and checked it against the page's stated outputs and guardrails.
Verified
Nov 13, 2026: source link resolves to the original, the credited creator published it, our excerpt or prompt matches it, and the license allows our use.

The original is licensed under Apache-2.0 by anthropics. Modified: our adapted prompt and summary are our own wording; the original skill {"was unchanged at the source when we snapshotted it (live source later removed)" if src_removed(s) else "is unchanged at the source link"}.

Are you the creator? Claim or correct this page.

More in Operations & Admin

All 203 Operations & Admin playbooks →

Related playbooks

Operations & AdminSkill✓ AI dry run

Compare design options with a weighted score and stress test

Use the trade-study-report skill from Anthropic's Claude for Manufacturing plugin to turn a filled-in options matrix and test results into a ranked, sourced comparison, set aside any option that breaks a hard requirement, vary the weights to see whether the winner changes, and draft a decision memo an engineer reviews and signs.

Credit: anthropics

Under an hour
Operations & AdminSkill✓ AI dry run

Connect your agent to a system with no ready connector

Use the custom-connector skill from Anthropic's Claude for Manufacturing plugin to name the exact system and one data flow, check for an existing vendor or community MCP server before building anything, settle on the highest workable access route down to a scheduled export, record why, keep credentials out of the chat and explain how the chosen route is likely to break.

Credit: anthropics

An afternoon+
Operations & AdminSkill✓ AI dry run

Diagnose a down machine from its history, sensors and manual

Use the breakdown-diagnosis skill from Anthropic's Claude for Manufacturing plugin to put any safety hazard first, read the sensor trace before the stoppage, search the asset's work orders and technician notes, check the manufacturer's troubleshooting guide, and get ranked likely causes with the quickest test to tell them apart plus a draft preventive maintenance change.

Credit: anthropics

An afternoon+
Operations & AdminPrompt✓ AI dry run

Schedule a four-week trust ramp instead of micromanaging

Move from reading every draft in week one to weekly summaries in week four, with an interrupt contract limited to approvals, missing data or out-of-scope work.

Credit: Carnage (@0xCarnagee)

Minutes