Skills
Skills
Skills are written instructions an agent loads only when it needs them, such as how to file an invoice or answer a support thread (module skills). Only a skill's one-line description sits in an agent's prompt; the rest is loaded when that line fits the job. A skill reaches agents only once a person publishes it.
In the sidebar under Agents → Skills and Agents → Skill hub.
Skills (Library)
The Library tab holds your organisation's skills.
- The skills list. Every skill with its title and description, its state (
draft,publishedorarchived), its version, and how often it has been loaded. Click one to read it in full, with Publish or Unpublish if you may. - Proposed by agents. Drafts agents proposed from their own work, and installs from the hub that agents proposed, waiting for a person.
- Which agents carry it. For a skill you open, each agent is listed with whether it carries it, how (On demand or Always: its description goes in the agent's prompt on every turn), and whether it came through a set.
- Where skills come from. Repositories skills are imported from, with when each was last synced, how it went, and whether a token is stored.
- When the hub has a newer version of a skill you installed, the skill says so; see the differences first.
An agent with no skills assigned sees every published skill. Once it is given one, it sees only what it was given.
Write a skill
- Press Write a skill.
- Give it a Name (how an agent asks for it, such as
release-notes), a Title and a Description. The description is all a run sees before deciding to load it. - Write the Instructions and press Create. It starts as a draft.
Publish or unpublish
Open the skill and press Publish or Unpublish. A published skill cannot be edited: unpublish it, edit the draft, and publish again. Each publish keeps a version you can restore.
Give a skill to an agent
Open a published skill, find the agent, and press On demand or Always. Take off removes it. A skill that came with a set is changed by editing the set.
Import skills from a repository
Under Where skills come from, press Add source, enter the Repository, optionally a Directory and an Access token (encrypted, never shown again). Press Preview to see what would come in, then Import. Imported skills arrive as drafts and never overwrite anything published.
Skill hub
The Skill hub gathers skills from GitHub repositories, awesome lists and pastes. Each is audited before anyone can install it, and shows its Audit, Risk, Licence, Source, Category and whether it Works in Max. Skills that do the same job are grouped as variants: Compare variants to pick one. Scripts come as text and are never run.
Install a skill from the hub
- Open Agents → Skill hub and find the skill.
- Press Install and choose Install as: Draft — read it first, or Published — agents can load it now.
- It is copied into your Library at that exact version. Copies already installed keep working if the hub changes.
For platform administrators
Platform administrators also keep the hub here:
- Sources: add a hub source (GitHub repository, Awesome list, Folder or Paste), how often it is refreshed (Daily, Weekly or Never), and Sync it now.
- Curate the selection: merge or split variants. A merge or split stays as you made it.
- Skill sets: curated sets every organisation can install, and a template can carry.
- Templates: ready-made agent and team templates, with their instructions, suggested modules, tools at hand, rule presets, schedules and models per job. Add starter templates adds the built-in ones. Every published version of a template is kept.
Who can do what
| Permission | Label | Lets you |
|---|---|---|
skills:read | View and load skills | Open both screens, read skills, their files, versions and use, and preview a source. |
skills:write | Write and edit skills | Write and edit drafts and their files, give skills to agents, make sets, install from the hub, and add, import and remove sources. |
skills:publish | Publish skills | Publish and unpublish skills. |
Writing and publishing are separate on purpose, so an agent cannot write its own standing orders and put them into use. Keeping the hub's sources and templates needs a platform administrator.
What happens on its own
| Job | When | What it does |
|---|---|---|
hub-sync | Every hour, at 17 past | Reads the hub sources that are due, groups what they brought into variants, and counts how many organisations installed each skill. A pass cut short by a limit carries on the next hour. |
reflect | Every day at 04:30 UTC | Reads the previous day's failed skill uses and proposes fixes as draft skills. Publishing stays with a person. |
Needs
Agents switched on.
What agents can do
An agent lists and loads published skills, can propose a new skill from what it just did, and can propose installs from the hub. A proposal is a draft until a person publishes it; a published skill is refused for editing. Attached scripts are stored, never run. Read-only tools run at once; the rest ask first under the default rules.
- Using skills, read-only:
skills.list,skills.load - The whole shelf, read-only:
skills.skills,skills.skill,skills.lint,skills.used,skills.versions,skills.sets,skills.carriedBy,skills.carries,skills.files,skills.file,skills.sources - Hub, read-only:
skills.hubSearch,skills.hubGroups(team templates),skills.hubTemplates(role templates) - Change:
skills.propose,skills.proposeInstall(installs nothing; waits for a person),skills.update,skills.addFile,skills.updateFile,skills.removeFile(deletes),skills.restoreVersion,skills.remove(deletes a skill that is not published),skills.unassign(takes a skill off an agent),skills.createSet(starts empty),skills.removeSource(deletes) - Reaches outside:
skills.previewSourcereads the repository as it is now, and asks first by default.
Over MCP the tools are named skills__list and so on.
Related
- Agents: who carries skills.
- Knowledge: what agents remember, as opposed to instructions they load.
- Agents and approvals