Skills

A skill is a Markdown playbook with optional supporting files. The agent sees each skill’s name and description, and reads the full text only when a task needs it. That keeps context small, which is why guidance that is only sometimes relevant belongs in a skill rather than an instruction.

Layout

• Directoryskills/

  • Directoryreview/

    • SKILL.md
    • Directoryreferences/

      • style.md
    • Directoryscripts/

      • check.py

Each directory under skills/ that contains a SKILL.md is one skill. Supporting files must be text. Binaries belong in the sandbox image or an environment.

SKILL.md

skills/release-notes/SKILL.md

---
name: release-notes
description: Draft release notes from the changes merged since the last release.
---

List the changes merged since the last release, group them by user-visible impact,
and write one plain sentence for each. Put breaking changes first.

Frontmatter

Field	Required	Meaning
name	No	Identifies the skill across updates. Defaults to the directory name. 1 to 128 ASCII letters, digits, dots, underscores or hyphens; starts with a letter or digit and cannot end with a dot.
description	No	One line telling the agent when to use the skill. If omitted, Cengrid derives it from the first prose line of the body, capped at 200 characters. Write it yourself.
requiredCapabilities	No	Capabilities the skill needs, such as egress:api.linear.app.
requiresConnectors	No	Account connectors the skill needs, such as google or linear.
declaredCreds, egress	No	Recognised by the importer. egress is read as a shorthand for requiredCapabilities.
private, agent_only, scope, visibility	No	Mark a skill as private to the agent or to a person. Do not use them in plugin skills. A plugin skill marked private, or scoped to personal, private or me, is rejected, which rejects the plugin.

A skill with a blank body is an error, and any error rejects the whole plugin on connect or re-sync.

Caution

Skills that set allowed-tools, context, agent, hooks or disable-model-invocation are imported as unavailable with the reason shown, rather than having those restrictions silently dropped.

A skill that uses a connected account

The starter google-workspace skill shows the pattern for a per-person OAuth connector:

skills/google-workspace/SKILL.md (abridged)

---
name: google-workspace
description: Read and act on the user's Gmail, Google Calendar, and Google Tasks through per-user OAuth.
requiredCapabilities:
  - egress:gmail.googleapis.com
  - egress:www.googleapis.com
requiresConnectors:
  - google
---

Use this skill when the user asks about Gmail, Google Calendar, or Google Tasks.

The resolved user's Google OAuth token is available as an environment variable,
one per API host: `$VAULT_TOKEN_GMAIL_GOOGLEAPIS_COM` for `gmail.googleapis.com`.
Do not ask the user for a token, log it, or use another principal's.

Who can create skills

• Administrators create custom skills in the console, at the organisation or a team.
• The agent creates skills in the bucket of the scope it is working in. In a team’s rooms, that is the team’s bucket.
• Plugin and pack skills are read-only in the console. Edit them in GitHub and Re-sync.

Skills can also be shared by grant, and promoted to the whole organisation by an administrator. Skill packs, the older form that imports skills from a git repository, are still accepted under Available.