Skills
Where do skills live?#
On the Skills card in Settings ▸ Prompts. Each skill shows where it runs and who it runs for, with a Running / Paused switch; Write a skill opens the editor.
Skills sit on your organization's shelf, the same one that holds the agent's memory, so a skill written on one project works for every project of that organization.
Write a skill#
The editor takes over the Settings page; Back returns you to the card.
Name is what the card shows — for example, "House style for API pages".
When should the agent reach for it? is the trigger, not the job: "when a page about an endpoint has no example request". It is the one field that decides whether the skill is ever used.
Instructions are the steps, in order, and what never to do — in Markdown.
Tick the doors under Where does it run? — see the table below.
Tick roles under Who does it run for?, or particular people under Or particular people. Nothing ticked means anybody.
Press Create skill. It starts running at once on the doors you ticked.
Where can a skill run?#
One agent answers through four doors, and a skill runs only behind the ones you tick:
| Door | What it is | Ticked for a new skill? |
|---|---|---|
| Public docs chat | The AI chat readers open on your published site | No |
| Admin chat | Your own chat in the panel | Yes |
| Public project MCP | The anonymous MCP server your readers' agents connect to | No |
| Admin MCP | Your MCP token, and the runs the Docsbook agent makes for you | Yes |
Nothing ticked means nowhere: the skill is saved and never runs. The two public doors change what readers see — the card says so next to them — so a skill reaches readers only when you tick one.
How a skill arrives depends on the door:
- In the two chats, its text goes into the agent's instructions on every turn — up to 8,000 characters of each skill and 24,000 in all. A skill past that total is still named, with its trigger, so the agent knows it exists.
- On the two MCP servers,
find_skillanswers with your skills first — name, trigger and where it runs — ahead of Docsbook's public catalog.
Who does a skill run for?#
Roles narrow a skill on any door: Readers (no account), Viewers, Editors, Admins / full access and Owner.
Or particular people lists everyone with access to the project, including invites not yet accepted. A restriction is met only by evidence:
- A visitor with no account counts as a reader, never as "anyone".
- A skill aimed at named people runs only where the agent knows the person's email — so never for an anonymous reader, and never over an MCP token.
Write one with the agent#
Ask in the admin chat — "Write a skill: every API page opens with a working request" — and the agent drafts it, saves it and tells you where it runs. New skills go on your two own doors; it asks before putting one on Public docs chat.
A good skill is short and specific:
Name: House style for API pages
When should the agent reach for it? When it writes or edits a page about an endpoint.
Instructions:
1. Open with one sentence: what the endpoint does and who calls it.
2. Show a working curl request before the parameter table.
3. List every error the endpoint returns, with what the caller should do.
4. Never state a limit or a default the source does not — link where it is stated, or leave it out.FAQ#
Who can change skills? Owners and admins — full access to the project. Everyone else who can open it sees which skills are in force, and the agent refuses to save a skill for them.
How is a skill different from Agent instructions? Agent instructions, on the same tab, are standing rules for your own agent on every turn and every run. A skill is one way of working with a trigger, and you choose its doors — including the ones your readers use.
Can a skill make the agent say something untrue? No. Skills rank above the agent's own habits, but not above its honesty rules: a skill cannot authorise inventing a fact, promising a capability or claiming work that was not done.