Skills
A skill is a procedure an agent follows: the steps in order, when to stop, and the shape the result has to take. An agent fetches one when the task in front of it matches. Skills live in Build → Skills.
A skill proposes. It never changes a record.
How it works
A skill has two boxes.
The procedure is markdown, and agents receive it word for word: the steps, the stop conditions — “if the scope exceeds one pass, say so and propose sub-scopes” — and what to read as data rather than as instruction.
The output shape says what the answer must look like: a review table, a findings list, a report. It stays in its own box because you cannot review free-form analysis. Two runs of the same skill come back the same shape, so you can compare them.
Write one
Click +. A new skill opens as a worked example you type over, not an empty box.
The page has three tabs: Procedure, which holds the output shape too, Meaning for the synonyms the catalog advertises, and Domains for who receives it. Identity sits behind the settings cog. Preview, Versions, Activate and Save sit in the header.
One Preview / Markdown control moves both boxes, since you read and write a procedure and its shape together. A skill opens ready to type. Switch to Preview to read it rendered, and the choice sticks. Clear empties both, behind a confirmation.
Search covers the procedure text, not just names, so a remembered step finds its skill.
Publish it
Activate with a comment publishes a numbered version. Agents follow the active version, so editing a draft changes nothing an agent runs until you activate. Activating an older version rolls back. Deactivate takes a skill out of service immediately. → Versions
Copy duplicates a procedure as an inactive, unrouted draft with its lineage stamped — the starting point for a regional or team variant.
Choose who receives it
Add a skill on a domain’s Skills tab: search, add, drag to order. That order is the order the domain’s catalog lists them in.
The catalog shows each one by name and a single line. Get Skill returns the full text when the task matches.
A skill in no domain reaches nobody. Domains decide who receives what.
See what an agent gets
Preview shows the served payload three ways: Formatted, Markdown — verbatim, what the agent receives — and JSON.
Notes
- This page lists the skills your org writes. The seven the package ships run the admin agent on its own topic and MCP server, which the permission model gates. Studio neither lists nor edits them. → The admin agent’s skills
- Seeded skills arrive with v1 active, so the shipped procedures work out of the box.