Skills

Skills are specialized instruction sets that load only when their trigger fires. Cold context stays small; hot context stays relevant.

Triggers

A skill has one or more triggers:

Keyword — e.g. architecture, react-router.
File pattern — e.g. *.tsx, Dockerfile.
Project signal — e.g. package.json#dependencies.react-router-dom.
Manual — the orchestrator or a human asks for it explicitly.

Skill layout

tree

.ai/skills/
  react-router/
    SKILL.md          # the prompt body
    skill.json        # id, triggers, scope, owner
    evals/            # at least one eval case

Authoring a skill

Create the folder under .ai/skills/<name>/.
Write skill.json with triggers and scope.
Write SKILL.md with the actual instructions.
Add at least one eval case in .ai/evals/cases/ that exercises the skill.
In the Alfred source repo, run cd packages/evals && npm run test.
Open a PR. The librarian reviews triggers; the qa reviews eval coverage.

Skills cannot override security policy. A skill may teach an agent how to ask for a destructive command, but the permission gate is owned by the kernel, not the skill.

Skill packs

A skill pack is a versioned bundle of related skills plus a shared eval suite. Alfred ships two packs:

@alfred/skills-core — the kernel, the panel, the policy.
@alfred/skills-web — TypeScript, React Router, TanStack, Node, Postgres.

Pack your own by keeping skills under .ai/skills/ and adding eval coverage before review.