Start with a reusable instruction.
The Cowork example is a small plugin with two Markdown commands. It asks the conversation model to follow a repeatable process; it does not draw a native pane.
- Download kitchen-cowork.zip and keep it zipped.
- In the Cowork or unified Claude desktop experience, use Customize → Plugins and the custom upload option. Availability and labels can vary by account and organization policy.
- Run
/kitchen-cowork:tour. - Attach the synthetic workshop notes, then run
/kitchen-cowork:experiment.
/kitchen-cowork:tour /kitchen-cowork:experiment
What should happen?
The command asks for four fields: Goal, Smallest change, Observable result, and Next question. Output can vary because the model follows instructions rather than a deterministic schema.
ILLUSTRATIVE OUTPUT SHAPE
- Goal
- Produce a consistent experiment card from the supplied notes.
- Smallest change
- Rename one output field in the command.
- Observable result
- The next card uses the new field name.
- Next question
- Which other part of the workflow is worth making repeatable?
Make the first edit.
In plugins/kitchen-cowork/commands/experiment.md, rename Next question to What to try next. From the repository root run npm run package:cowork, then upload the new ZIP and try again with the same notes. Packaging needs Python 3.9+; the npm wrapper also needs Node.
Invoking these commands uses normal conversation-model usage. If the session cannot access the bundled notes, attach them explicitly.