Claude for Manufacturing is a plugin that gives Claude written instructions for manufacturing tasks. It is not a new model, and it does not change how you get at your data. That is worth knowing before your LinkedIn feed convinces you otherwise.
Anthropic launched it on 12 November 2026, and within hours the “bombshell” posts arrived. We like the attention manufacturing is getting. But we wanted to know what is actually inside, so we installed it and ran one of its skills on our own simulated dam.
Key Takeaways
- Claude for Manufacturing is a plugin with 17 skills. A skill is a text file telling Claude how to do a specific task.
- It doesn’t change the models behind Claude, data quality, or the need for data context.
- It doesn’t operate equipment. It drafts, analyses and ranks; people decide.
- In our test, Anthropic’s root cause analysis (RCA) skill and a simple skill we wrote ourselves gave very similar results on the same data.
- The hard part is still the data: access, alignment, quality, and the links between data sets. No skill does that for you.
Watch: Claude for Manufacturing, Tested on a Simulated Dam
What is Claude for Manufacturing?
It is a plugin for Claude, available on Team and Enterprise plans. It bundles 17 preconfigured skills for roles such as estimators, quality engineers, technicians and buyers. The skills cover areas like quoting, quality investigations, breakdown diagnosis, purchasing and production reporting. You can find Anthropic’s own description on claude.com.
Anthropic has been clear about the limits. The plugin doesn’t operate equipment, and where it proposes actions, a person approves them.
Get our weekly free newsletter where we bundle our insights in Industrial AI and DataOps
What is a Skill?
A skill is a set of instructions written in plain text. It describes the inputs, the workflow, and the format of the output. When Claude gets a task that matches, it follows those instructions instead of improvising.
That is the whole mechanism. If you have ever written down how an experienced engineer approaches a breakdown, you have written something very close to a skill. (We did exactly that last week, for our own training material. More on that below.)
This is also why we’d advise reading the skills rather than just using them. They show what Anthropic believes the manufacturing domain is about, and you can borrow, edit and improve them for your own plant.
What We Tested: the Root Cause Analysis Skill
An RCA builds a timeline of relevant events, writes down hypotheses about what could have caused the failure, and then validates those hypotheses against evidence. It is slow, detail-heavy work, and a large language model can help with it.
To test Anthropic’s RCA skill, we used the fictional Copper Canyon Dam from our Virtual Maintenance Engineer demo. It’s a fully simulated dam with three turbines. In our scenario, Turbine 1 was stopped after a high-high temperature alarm on a guide bearing pad, and the shift followed the written procedure.
The data set is what you would expect to find in a real plant:
- Historian exports (a three-week trend)
- Shift logs and work orders
- Oil analysis and reliability data
- A previous RCA
- Internal documents and supplier documents
We ran the same prompt, on the same data, twice, both times in a terminal with Claude Code. The first run used the skill from Anthropic’s plugin. The second used a short skill we had written ourselves, which describes the available data, the working procedure and the output location.
Each run took five to seven minutes. Each produced a markdown report with a timeline, collected evidence, hypothesis checks and ranked candidate causes. Both pointed to the same area: bearing overheating on the guide bearing pad, with oil condition and pad design as contributing factors. The two reports were very similar.
We won’t pretend this is a benchmark. It is one scenario, one data set, one run each, and we scored nothing. What it does show is that a well-structured instruction file, whether from Anthropic or from you, takes you a long way. We’d rather you check this yourself than take our word for it.
Where is the Real Difficulty?
Not in the skill. The data we used was generated for the demo, and it was prepared by hand. In a real plant, the work that decides whether this approach succeeds is everything that happens before the prompt:
- Access. Can you actually get to the historian, the CMMS, the shift logs and the lab results?
- Alignment. Do tags, assets, timestamps and work orders refer to the same thing across systems?
- Quality. Are the sensor values and the free-text entries trustworthy?
- Context. Are the links between data sets there, so that a vibration trend can be tied to an asset, a work order and an oil sample?
Want to learn more about Industrial Data Architectures? We have a book and an academy.
How to Try It Yourself
We have published everything for free. No secrets, because we need to learn from each other. Pick the route that fits your skills:
- Starter pack (no installation). Download the zip from the Virtual Maintenance Engineer page, upload the files to any chatbot, and follow the README. Free versions work, and it doesn’t have to be Claude.
- RCA bot pack (terminal). Run the same exercise with an agentic tool such as Claude Code. This is where you can load Anthropic’s skill, write your own, and compare the two as we did. You’ll need some technical comfort, for example Python.
- Full environment (GitHub). Clone our Cromley Hydraulics repository, run
docker compose up, and you have the dam, its SCADA and an instructor console where you can inject faults yourself. We also have a chemical plant and a bakery you can clone.
Whichever route you take, add the plugin to your environment and read the skills. Then run the same prompt with and without them, and compare the output.
One caution: don’t paste company data into public AI services. Use the tools your company provides and allows. Someone in your organisation will care about that, and rightly so.
Frequently Asked Questions
Is Claude for Manufacturing a new AI model?
No. It is a plugin that adds skills to the existing Claude models. The models, and what they can do with your data, stay the same.
Can it control machines or run a production line?
No. It drafts documents, analyses records and ranks likely causes for people to review. It doesn’t operate equipment.
Do I need Claude for Manufacturing to run an AI-assisted root cause analysis?
Not necessarily. In our test, a short skill we wrote ourselves produced a very similar result. The plugin is a useful starting point, and a good source of ideas for your own instructions.
Does it work for process manufacturing too?
Our test used a dam with turbines, which is asset-heavy rather than discrete. The skills were written with specific tasks in mind, such as quoting and quality paperwork, so check them against your own processes before relying on them.
Is this safe to use with my plant’s data?
That depends on your company’s rules and the plan you use. Start with simulated or public data, as in our demo, and involve your IT and security teams before connecting real systems.