The crew

Skills

A skill is a folder with a SKILL.md inside. Drop one into your workspace and Brigade discovers it next turn. Skills are prompt-resident: the name and description are injected into the system prompt, and the agent reads the body on demand when the skill is relevant. 56 skills ship bundled.

Anatomy of a skill#

A skill directory has a SKILL.md with frontmatter. The name and description are the discoverability hook — the description is what the model sees when deciding whether to open the skill. Optional eligibility metadata gates when it loads.

skills/release-notes/SKILL.md
---name: release-notesdescription: Draft release notes from a range of git commits.--- # Release notes Steps the agent should follow when asked to draft release notes…

Six discovery roots#

Skills merge from six roots in precedence order (later roots win on a name clash), so a workspace skill can override a bundled one:

#RootUse
1<install>/skills/Bundled skills shipped with Brigade
2config.skills.pathsExtra paths from brigade.json
3~/.brigade/skills/Managed via manage_skill
4~/.agents/skills/Personal, across projects
5<workspace>/.agents/skills/Project-scoped
6<workspace>/skills/Workspace-scoped (highest precedence)

When an org is configured, a seventh source is merged just above the bundled root: skills shared by org-visible peer agents, so a crew can pool skills along its reporting lines.

Eligibility filters#

A skill only loads when it is eligible for the current environment and agent. Filters cover:

  • Operating system.
  • Required binaries on PATH.
  • Required environment variables.
  • Config-disabled state, and the per-agent skills allowlist.

A per-agent noSkills seal prevents the current working directory and .pi from leaking into the skill space.

Managing skills#

Inspect what is installed from the CLI:

terminal
$ brigade skills list$ brigade skills info release-notes

From inside a turn, the owner-only manage_skill tool manages skills against a specific agent or the managed (~/.brigade/skills/) root. Its actions are create, patch (append to a SKILL.md), delete, list, write_file (add a support file), and remove_file.

Why prompt-resident?

Injecting only the name and description keeps the prompt small while still letting the model choose the right skill. The full instructions load only when the skill is actually used — so you can ship dozens of skills without bloating every turn.