The full power of TDengine, now free forever for up to 5,000 tags.

Explore More

How Does Claude Breakdown-Diagnosis Investigate a Machine Failure?

Juno Qiu

November 18, 2026 /

The breakdown-diagnosis Skill organizes evidence for a machine failure and drafts ranked possible causes with tests that could distinguish them. It uses maintenance history and technician close-out notes, with sensor history, photographs and the manufacturer’s troubleshooting information when available. Its output is advisory: the technician decides what to inspect and what work to perform.

Anthropic’s Skill instructions place safety implications before diagnosis and describe operation even without sensor data or a connector. A thinner evidence set should produce a correspondingly limited result.

Worked example: investigate a pump stop

This example is fictional. The prompt and review steps follow the published Skill instructions; they are not a recorded Claude session or maintenance recommendation.

A technician reports that pump P-104 stopped during a production run. Prepare four pieces of evidence:

Input Example to supply Check before use
Event Operator says the stop was approximately 10:15 local time Is this the observed stop or the later ticket-entry time?
CMMS Asset ID, previous tickets and technician close-out notes Do the work orders refer to this physical pump?
Sensor export Bearing temperature and operating state from before and after the reported stop Include timestamps, time zone, units and coverage; check clock alignment
Manual or photo Relevant troubleshooting page and an approved photo Confirm the manual applies to the installed model

After making the files available to Claude, ask:

Investigate the reported stop of pump P-104 at about 10:15 local time.
Use the attached maintenance history, sensor export and manual excerpt.
State safety concerns first. Separate observations from hypotheses, cite
specific work orders, list missing evidence, and propose a safe test that
would distinguish the leading candidates. Do not change a work order or PM plan.

Inspect the response in this order. First, can you trace each quoted observation and work-order reference to the supplied material? Second, does the sensor window actually cover the uncertain event time? Third, does each hypothesis include evidence against it as well as support? Finally, can a qualified technician review the proposed test under the site’s safety procedure?

An illustrative review could consider bearing wear, a cooling problem and an unreliable sensor as competing explanations for a high temperature. If no inspection or independent measurement is available, none is established. A suggestion to shorten a maintenance interval is a draft for the maintenance owner, not an implemented change.

For more on connecting a measurement to the correct equipment and operating context, see industrial data contextualization.

Give the Skill an identifiable event

Provide the asset and site, observed symptoms, approximate failure time and operating state. Add the relevant CMMS work orders, including actual close-out notes rather than only failure codes. Supply the manual’s troubleshooting section and photos where appropriate. If sensor data is available, identify the signals, units and a bounded pre-failure window, plus the timestamps and source.

If the only time available is when the work order was opened, say so. That time might be later than the failure and can change which sensor samples are relevant. Confirm that asset identifiers in the CMMS and historian refer to the same physical machine.

Ask Claude to diagnose the named breakdown and describe the evidence you supplied. Review whether it distinguishes the observation, the source record and its inference. The Skill is intended to cite prior work orders and suggest a low-cost discriminating test for each candidate. Check the actual records before accepting a citation.

Review a candidate, not just the ranking

Suppose a pump stopped and the records show a high temperature before the reported stop. A bearing problem may be one candidate, but a blocked cooling path or a faulty temperature measurement could also explain part of the observation. Ask what evidence supports each, what contradicts it, and which safe inspection would separate them. This is an illustrative example, not a finding from a real machine.

The Skill may draft a preventive-maintenance interval change and produce a downtime record for a subsequent report. A draft does not establish that the change is appropriate, approved, or entered into the CMMS. Where the breakdown-to-next-report chain is used, the diagnosis remains advisory and is not written back to a ticket according to that chain’s instructions.

Work within the evidence and safety limits

A missing sensor window does not prevent a history-based draft, but it removes evidence that might distinguish candidates. An outdated manual, ambiguous asset mapping or unverified timestamps should be visible in the result. Follow the plant’s isolation and safety procedures; do not execute a suggested test until a qualified person has checked its suitability.

Related Articles