agent-skills

@mikestopcontinues/agent-skills

Lifecycle skills, persona agents, and governance hooks for autonomous-project workflows — authored once in TypeScript, shipped as native plugin bundles for Claude Code, Codex, and OpenCode (plus universal skills install via vercel-labs/skills).

The repo IS the deliverable: clone (or point a marketplace at the GitHub URL) and any of four install verbs picks up the right bundle.

What you get

Component	Count	Examples
Skills	21	create-spike, create-plan, execute-plan, validate-doc, triage-feedback, process-feedback, lock-decisions, start-session, yolo-project, …
Persona agents	11	architect, qa, dx, reviewer, security, implementer, designer, researcher, verifier, doc-writer, conflict-resolver
Governance hooks	5	block-main-edits (refuse git commit on main/master), pre-commit-gate (run pnpm run check), pre-write-doc (require yolo new scaffolding for new docs), pre-bash-doc (verify doc-link integrity), verify-worktree-path (refuse edits outside the worktree)
CLI scripts	2	yolo new <kind> <args> (project/spike/plan/chapter/review/note/decision scaffolder), doc-check-links.sh (cross-reference checker)

The skill / agent / hook inventory is canonical at generator/canonical-sources/. Each harness emit ships the same content in its native layout.

Install

Pick your harness. The first three install verbs all read this repo; the OpenCode npm path resolves the same content via package.json#main.

Claude Code

claude plugin marketplace add github:mikestopcontinues/agent-skills
claude plugin install agent-skills@agent-skills-dev

That's it. Skills appear as @mikestopcontinues/agent-skills:<skill-name>, the 11 personas register as dispatchable subagent types, and the 5 hooks fire automatically on matching tool events (no per-hook trust step — CC treats plugin-bundled hooks as managed).

Codex

# 1. Add the marketplace and enable the plugin
codex plugin marketplace add mikestopcontinues/agent-skills
# Then either toggle on inside the TUI (/plugins → Agent Skills → Install)
# or add to ~/.codex/config.toml manually:
#   [plugins."agent-skills@agent-skills-dev"]
#   enabled = true

# 2. Enable the plugin_hooks feature flag
# Required for any plugin-bundled hooks to be discovered (Codex ≤0.130 keeps
# this behind a flag). The 21 skills work without it; the 5 hooks do not.
codex features enable plugin_hooks

# 3. Trust the 5 plugin hooks (one-time, interactive)
# Open the TUI: /plugins → Agent Skills → Hooks
# Toggle each hook (block-main-edits, pre-bash-doc, pre-commit-gate,
# pre-write-doc, verify-worktree-path) to Trusted. Codex writes
# [hooks.state."agent-skills@agent-skills-dev:hooks.json:..."] trusted_hash
# entries to ~/.codex/config.toml. Hooks fire on subsequent `codex exec` /
# `codex` sessions.
codex

Skills appear as agent-skills:<skill-name>. The persona agents ship as TOML files under codex/agents/ and are loaded by the same plugin install.

Universal (vercel-labs/skills) — skills only

# Per-project (installs to ./.agents/skills/<name>/)
npx skills add mikestopcontinues/agent-skills

# User-wide (installs to ~/.agents/skills/<name>/)
npx skills add -g mikestopcontinues/agent-skills

Works for [47 agent products][vercel-labs/skills] including Claude Code, Codex, OpenCode, Cursor, Gemini CLI, Warp, GitHub Copilot. Installs the skills only — no persona agents, no hooks, no scripts — because the universal .agents/skills/<name>/SKILL.md convention is the only thing all 47 agents agree on. For full hook / agent coverage on a specific harness, use that harness's native install path above.

OpenCode

OpenCode has two install paths because its plugin and skill systems are separate concerns:

# Skills (via the universal route)
npx skills add mikestopcontinues/agent-skills

# Hooks (via npm + opencode.json)
npm install @mikestopcontinues/agent-skills@github:mikestopcontinues/agent-skills

Then in opencode.json:

{
  "plugin": ["@mikestopcontinues/agent-skills"]
}

The npm package's package.json#main points at ./opencode/src/plugin.ts which OpenCode loads to register the 5 hooks via the tool.execute.before plugin API. Skills come from the .agents/skills/ install path that OpenCode reads natively.

What's in here