Using Skills

using-skills

Route by intent. Every name below is an entry in skills-lock.json, the installed roster, so a name here always resolves on this machine.

The kernel routing index in ~/.claude/CLAUDE.md binds first. When the kernel names a trigger — delegation, done/stuck claims, unclear acceptance, retries, loop-shaped work, code changes, output shape — its route wins. A few of those skills still get a row below so that every lock name resolves in one place; the row is a lookup, not a second opinion.

Two hops maximum: using-skills → domain router → member. Never route back here from inside a domain router. Read the live SKILL.md before acting; never act from memory of an old roster.

Mode column: Skill() = model-invocable · manual = read its SKILL.md inline or use its slash command · via-router = enter through the router named.

Hop 1: domain routers

Shape an idea before building

Write and change code

Diagnose and verify

Delegate and orchestrate

Read the outside world

Write for people to read

Visual and HTML output (all through using-design-skills)

Listed so a name resolves, not as a bypass. The router picks the owner.

Git, releases, dependencies

Session and fleet upkeep

Repo-specific

pierre-guard (guards the @pierre/diffs integration) and release-plannotator (release notes, version bumps) apply inside the Plannotator repo only.

Not in this map

Plugin skills live with their plugins, not in skills-lock.json: code-review, code-simplifier, commit-commands, frontend-design, figma, session-report, security-guidance, and the rest under ~/.claude/plugins/cache/. The active available-skills listing shows them. Rules files (~/.agents/rules/*.md) are not skills; the kernel routes them.

Subagent exemption

A worker executing one assigned task does not enter this router. The dispatcher already routed; the worker follows its brief.

Red flags

Naming one and proceeding anyway needs a stated reason.

• 「這只是小問題，不用 skill」: go direct when the domain is obvious, but do not skip routing on substantial work.
• 「我記得那支 skill 的內容」: members change. Read the live SKILL.md first.
• 「先做完再回頭套流程」: for loop-shaped or visual work the router picks the process first. Retrofitting ships half-done work.

Freshness self-check

This map matches skills-lock.json at its last edit. When a name here does not resolve, or the listing shows a skill this map omits, confirm before trusting it:

diff <(rg -o '`[a-z][a-z0-9-]+`' ~/.agents/skills/using-skills/SKILL.md | tr -d '`' | sort -u) \
     <(jq -r '.skills|keys[]' ~/.agents/.skill-lock.json | sort)

Names in the left column only are stale references or prose. Names in the right column only are installed skills this map has not placed yet.

Nothing fits

Not loop-shaped, not visual, no intent above: check the active available-skills listing, then work inline. A new recurring intent goes to the user as a proposal for a skill or recipe. Never improvise a router inline.