Automation in a company brain is not scripts you write. It is skills: saved procedures your AI runs on demand, each one a repeatable task written down once so it runs the same way every time. Your AI discovers them with list_skills and loads one with get_skill, and every skill runs inside your session, so it only ever touches what you are allowed to touch. Some skills you write yourself for your own recurring tasks. Others come as a standing catalog that does the real maintenance work: blueprint-scan maps where the company knowledge lives at the start of a build, source-sweep keeps the brain current by re-walking your connected folders and updating the pages that changed, weekly-review compiles what changed in the week, change-review gives owners oversight of recent edits, and query-external-system routes a question to a connected tool through its card in the Systems Registry. The pattern under all of it is the one that runs through these docs: the AI does the work, you review and decide. Start by invoking a catalog skill by name and reading what it produced. Once you trust its output, make it a habit, weekly for the review, per department for the sweep. You are not wiring up infrastructure. You are turning on procedures that already know how to run.
Steps
See what skills the brain already has
Ask your AI to list the skills in your brain (list_skills). You will see your own saved procedures alongside the standing catalog: blueprint-scan, source-sweep, weekly-review, change-review, query-external-system. Each has a one-line description of when to use it and a status that says whether it is active.
Invoke one by name and read the result
Pick a catalog skill and ask for it in plain language, "run the weekly review" or "sweep the marketing folders". The AI loads the skill (get_skill) and follows it. The first time, read what it produced end to end. That is how you learn what the skill does and where its output needs a nudge.
Turn your own repeated prompts into skills
When you notice yourself typing a similar prompt twice (summarise this transcript into our format, draft the update from this week), ask the AI to save it as a skill. It becomes a named procedure anyone with access can invoke. You edit it like any page, and there is no deploy step.
Make the useful ones a rhythm
The skills that earn their keep are the ones you run on a cadence: weekly-review every week, source-sweep per department, change-review when you sit down to approve recent edits. You do not have to remember every step, only to ask. The skill carries the procedure.
Keep the catalog trustworthy
A skill tells the AI what to do, so treat the list like a set of standing instructions. Turn off skills you do not use, read a third-party skill in full before setting it active, and review what changed when one is updated. A short list of trusted, active skills beats a long list of forgotten ones.
Resources
Troubleshooting
- A skill did something I did not expect.
- Ask the AI to show you the skill (get_skill) and read what it actually instructs. Skills are plain text, so adjust the wording or turn it off. Any change it made to the brain is reversible from the page History.
- I cannot tell which skills are active.
- Ask the AI to list skills with their status. Deactivate the ones you do not recognise or no longer use.
- A member ran a skill but it saw less than I did.
- That is by design. A skill runs in the caller session and is bounded by their access, so an owner run covers more of the brain than a member run. Nothing to fix.