Skills

Skill commands answer what capabilities an agent can use locally.

Two sources of skills

Skills reach a harness from two places, and the difference is the public/private boundary:

• The my-cli self-skill ships inside the CLI binary. It is organization-neutral — it only teaches harnesses how to drive my itself — so it is safe to install anywhere and carries no company content. The binary owns its lifecycle (install.sh, my setup, my ai, and a quiet refresh on human CLI runs, including after my update); manage it explicitly with my skills self ....
• Organization skills are declared in an org's manifest repo and land only once you my manifests add / my manifests sync that manifest. Because they live in a repo you control — typically private — they can carry organization-specific guidance. my ai composes them into the launch root's .agents/skills tree for harnesses with a project-local skill seam, with harness mirrors where needed. OpenCode is currently compatibility-global: present or explicit OpenCode setup/launch keeps org skills in ~/.config/opencode/skills, and my ai opencode --skills/--profile is rejected until a project-local seam is proven. my skills install / my skills sync remain explicit manual commands for user-global materializations.

Nothing organization-specific is baked into the public CLI: the self-skill stays generic, and everything particular to your organization lives in a manifest you own.

Launch-scoped skill selection

my ai composes the organization skill loadout for a launch and materializes it into the launch root (.agents/skills/, plus a .claude/skills/ mirror for Claude Code). Choose the loadout with mutually exclusive selectors:

my ai                                       # default loadout for this launch target
my ai --skills all                          # every declared organization skill
my ai --skills none                         # no organization skills (self-skill only)
my ai --skills acme:handbook,acme:support   # an explicit set of skill ids
my ai --profile support                     # a named loadout from the manifest `profiles` list

A profile is a named skill loadout declared in the manifest's profiles list, distinct from a role. With no selector, my ai uses the selected role's skills for a base umbrella launch, includes workspace-satisfied skills for session launches, uses all org skills for an unscoped umbrella, and intentionally uses no org skills for repo launches. These selectors compose for harnesses with a project-local skill seam (Claude Code, Codex, Antigravity). OpenCode is compatibility-global: it keeps organization skills in ~/.config/opencode/skills and rejects --skills/--profile until a launch-root seam is proven. The my skills commands below manage user-global materializations manually and are no longer the automatic setup path.

My AI self-skill

my skills self status --all
my skills self install --all
my skills self uninstall codex

The my-cli self-skill is bundled with the CLI and teaches harnesses how to use My AI itself. install.sh installs it into existing harnesses, my setup refreshes it with the selected harnesses, my ai ensures it exists for the selected filesystem harness before launch, and normal human CLI runs quietly align already-installed file-based copies with the running binary.

Inspect declared skills

my skills list
my skills show acme:handbook
my skills status

Use --json when an agent needs machine-readable output. my skills show also surfaces manifest requires entries such as service:<id>, which name services the skill expects the workspace to provide.

Install and reconcile

my skills install --all
my skills install codex --skill acme:handbook
my skills sync --all
my skills purge --all

Manifest install, uninstall, sync, and purge operate on explicit local harness materializations. They do not edit the manifest and are no longer the automatic setup path for organization skills.

sync installs or updates declared skills and prunes stale My AI-managed manifest targets by default. It leaves the bundled my-cli self-skill to my skills self .... Only the canonical id my:self is protected from pruning; a manifest-declared skill that happens to be named my is ordinary and removable. Pass --no-prune to only install or update. When the manifest itself changes, my manifests sync refreshes generated guidance, umbrella MCP config, and launch-scoped skill reconciliation notices for an existing matching umbrella unless --no-derived is passed.

Provenance

my records what it installed. It refuses to clobber a directory it did not place unless the command explicitly allows replacement.

Manifest skill directories are immutable source, not working storage. A skill that needs credentials, cookies, caches, downloads, or other mutable state uses ~/.local/state/my-cli/skills/<manifest>/<install-slug>/. my doctor warns when unexpected files appear under any declared skill source, and publish/sync holds them with unadopted_skill_source unless a reviewed source change was explicitly staged. This prevents runtime state from leaking into a manifest commit.

By default, filesystem harnesses receive symlinks. Use --copy to vendor a copy into the harness skill directory.

Supported harnesses

Harness	Install path
Claude Code	~/.claude/skills/<skill>
Codex	~/.codex/skills/<skill>
OpenCode	~/.config/opencode/skills/<skill>
Antigravity	~/.agents/skills/<skill>

Managed launches use those paths differently from manual skill commands: Claude Code receives a launch-root .claude/skills mirror, Codex and Antigravity read launch-root .agents/skills, and OpenCode stays global as a compatibility exception.