skills

A public collection of agent skills for Claude Code, Codex, and OpenCode, delivered with Vercel's skills CLI.

Each skill is a folder skills/<name>/SKILL.md with YAML frontmatter (name + description). Target Claude Code plus Codex when installing the full catalog: the rolling skills@latest CLI writes the canonical store under ~/.agents/skills/<name>/, which Codex and OpenCode read directly, and symlinks ~/.claude/skills/<name>/ for Claude Code. Restart the target agent if a newly installed skill does not appear.

Three options, pick one

🧰 Just agent-kit

The Claude Code-native agent team plus the full PRD lifecycle (11 skills).

npx -y skills@latest add vtmocanu/skills/skills/agent-kit -a claude-code -g -y

For automatic updates, first install the refresh helper with consumers of the skill store closed:

npx -y skills@latest add vtmocanu/skills --skill skill-maker -a claude-code codex -g -y

Claude Code auto-update hook, macOS/Linux only (requires Python 3 with fcntl; add to ~/.claude/settings.json):

"hooks": {
  "SessionStart": [
    {
      "matcher": "startup",
      "hooks": [
        {
          "type": "command",
          "command": "~/.agents/skills/skill-maker/scripts/refresh_skills.py --source vtmocanu/skills/skills/agent-kit",
          "async": true,
          "timeout": 300
        }
      ]
    }
  ]
}

Skill	What it does
agent-team	Auto-generate and run a per-repo Claude Code agent team: probe the repo, write .claude/agents/{role}.md subagent definitions from a role library, then orchestrate tasks with spawned teammates.
prd-create	Create documentation-first PRDs that guide development through user-facing content.
prd-start	Start working on a PRD implementation.
prd-next	Analyze a PRD and recommend the single highest-priority task to work on next.
prd-update-progress	Update PRD progress from git commits and code changes, enhanced by conversation context.
prd-update-decisions	Update a PRD from design decisions and strategic changes made during conversations.
prd-done	Complete a PRD: create branch, push changes, open a PR, merge, and close the issue.
prd-full	Run a PRD end-to-end autonomously (start, iterate until done, then PR), stopping after PR creation for review.
prd-close	Close a PRD that is already implemented or no longer needed.
prd-worktree	Create a git worktree for PRD work with a descriptive branch name (bundles a create.sh).
prds-get	Fetch all open GitHub issues in this project labeled PRD.

Anything later added under skills/agent-kit/ joins the bundle automatically.

📝 Just the PRD skills

The 10 prd-* skills from agent-kit, without agent-team. They work standalone; prd-create adds an agent-team architect reviewer only when the repo already has one.

npx -y skills@latest add vtmocanu/skills/skills/agent-kit/prd -a claude-code -g -y

For automatic updates, install the refresh helper and hook as in Just agent-kit, with --source vtmocanu/skills/skills/agent-kit/prd in the hook command.

📦 All skills

The whole catalog (24 skills), agent-kit included.

npx -y skills@latest add https://github.com/vtmocanu/skills -a claude-code codex -g -y

Claude Code auto-update hook, macOS/Linux only (requires Python 3 with fcntl; add to ~/.claude/settings.json):

"hooks": {
  "SessionStart": [
    {
      "matcher": "startup",
      "hooks": [
        {
          "type": "command",
          "command": "~/.agents/skills/skill-maker/scripts/refresh_skills.py",
          "async": true,
          "timeout": 300
        }
      ]
    }
  ]
}

Skill	What it does
away-mode	Let a session work while you are away (overnight, a weekend): it restates the grant, pre-asks the predictable questions, watches on a timer, decides reversible in-scope items with a buddy peer (or a subagent buddy when none is bound), holds irreversible, out-of-scope and denied actions, and greets you on return with one report of what was done, what was held, and the decisions left for you.
cicd-expert	The CI/CD expert. Generate a repo's pipelines through an interactive analyze-then-confirm conversation, or review, harden, debug, and speed up existing CI (supply-chain security, caching, path filters, job DAG, Renovate tool pins).
claude-permissions	Manage Claude Code permissions via Dippy (Bash/MCP allow/ask/deny plus the auto-mode [ASK] fallback wrapper, bundled) and settings.json (Read/WebFetch/Skill).
clodex	Runbook for clodex, which runs ChatGPT/Codex-plan and other OpenAI-compatible models inside Claude Code: install and upgrade, sign in, add or remove a model (/clodex <model-id>), re-patch after updates. User-invoked only.
done	End-of-session wrap-up: check git state across the directories touched this session, review for loose ends, and give a plain verdict on whether the session can be closed.
generate-dockerfile	Generate a production-ready, secure, multi-stage Dockerfile and .dockerignore for the project.
session-peers	Make Claude Code sessions and Codex CLI threads on one machine message each other. Async send supports handoffs; correlated ask/reply returns multi-round peer work to the current Codex turn without a stale queued reply; wait observes peer state. Codex threads can auto-attach as real peers and aliases follow renames. Auto-attachment requires running ~/.agents/skills/session-peers/scripts/peers.py install-hook --auto-attach once, then trusting the entry through Codex /hooks; see setup.
skipper	Coordinate enrolled agent leads across repositories, route user questions, track Claude context and handovers, and coordinate shared-resource holds. Technical work stays with leads and buddies.
reflect	Analyze the current session and propose improvements to the whole authored skill package, including instructions, scripts, tests, references, templates, assets, and hooks, then test and publish the approved changes.
skill-maker	Author, lint, and publish portable Claude Code, Codex, and OpenCode skills: canonical .agents/skills repo layout, Claude compatibility symlinks, cross-agent refresh hooks, agnix linting, and npx distribution scopes.
status-check	Report what is still open in the session (background jobs, items waiting on you, on another session, on CI or a bot, held items) as one table, once. /status-check every [N] repeats it on a session-only timer until nothing is left; /status-check stop ends the timer. Read-only.
token-audit	Audit a Claude Code setup for token waste (report only, change nothing): measure in-scope CLAUDE.md sizes and @imports, MCP servers/tool counts and whether tool deferral is active, any proxy that silently disables it, model/effort and mid-session switches, output-reducing hooks, per-agent model inheritance, cron/loop intervals vs the measured cache TTL, and the newest session log's cache-read/creation/input/output token split; emit one severity-ranked table plus the single highest-leverage fix.
upgrade-advisor	Evaluate a tool, framework, or dependency upgrade: discover the pinned version, find the latest installable one, read the changelog across the whole delta, and report which breaking changes actually touch this codebase (by grepping usage), with a safe / blocked / needs-work verdict.

Plus the 11 agent-kit skills from the table above.

Notes for all three options

• Refresh safety: Vercel installs into temporary projects; the wrapper publishes complete files with atomic replacement. Existing paths stay readable, unchanged files are untouched, and failed staging leaves the live installation intact. Retired supporting files remain available to already-loaded skills. This is atomic per file, not a frozen version of the whole skill. See refresh safety for the verified race, metadata compatibility, and cleanup limits.
• The wrapper uses add … --skill '*' in staging to discover new catalog skills. It also refreshes other tracked global skills and current-project dependencies by their recorded names and refs. It serializes cooperating refresh jobs and leaves repo-authored skills outside skills-lock.json untouched. Automatic refresh is optional; for manual use with agents running, invoke the same wrapper. Run direct modifying npx skills commands only with consumers of that store closed.
• The full-catalog installer targets claude-code codex explicitly. Codex supplies the universal .agents/skills destination, Claude receives a symlink to it, and OpenCode reads that same universal destination without another projection. A Claude-only target makes the CLI use copy mode and does not populate .agents/skills.
• Codex auto-update: point the Codex SessionStart handler at the same installed refresher, using ~/.codex/hooks.json with matcher startup|resume, then review and trust it through /hooks. For agent-kit-only or PRD-only installs, use the same --source argument as the Claude hook. Use the installed script or a thin machine-local trampoline that executes it; a detached copy of its implementation cannot receive updates. The hooks are asynchronous and best-effort: a session may begin with an older inventory, an early session end may cancel refresh, and the next start retries. The skill-maker hook section has the JSON contract and multi-source options.
• Renames and removals are not auto-pruned in a non-TTY hook; drop an old name with npx -y skills@latest remove <old> -g -y.
• Drop -g to install into the current project only (.agents/skills store plus agent-specific projections such as .claude/skills). Add -l to list without installing, or -s a b (space-separated) to pick a subset.
• Edit the source repo, never an npx-installed .agents/skills store or .claude/skills projection, which add/update overwrite.
• Agent definitions are different from skills. There is no portable .agents/agents directory; Claude Code, Codex, and OpenCode require separate native agent-definition files. See skill-maker.

Other files

• CLAUDE.example.md: a generic starter for ~/.claude/CLAUDE.md (Claude Code's global instructions), general AI-collaboration guidance only, no setup specifics. Not a skill; copy what's useful into your own config.
• retired/: skills no longer maintained or installed, kept for reference. See retired/README.md.

Contributing

Issues and PRs welcome. See CONTRIBUTING, the Code of Conduct, and the Security Policy.

Credits

The prd-*, cicd-expert, and generate-dockerfile skills draw on work vendored from vfarcic/dot-ai, created by Viktor Farcic and used under the MIT License (Copyright (c) 2025 Viktor Farcic). Most come from its shared-prompts/ directory; prd-worktree comes from .claude/skills/dot-ai-worktree-prd/ (renamed from worktree-prd, with its bundled create.sh). The prd-* and generate-dockerfile skills are copied largely verbatim, converted to the folder SKILL.md layout with the dot-ai category frontmatter dropped; cicd-expert keeps Viktor's interactive generator (formerly the generate-cicd skill) and substantially expands it with a security and speed reference set. Each keeps a provenance line pointing back to its source. Thank you to Viktor for the excellent PRD workflow and project generators.

The agent-team skill's initial design was based on Viktor's dot-agent-deck.

License

MIT © Vlad Mocanu. The prd-* and generate-dockerfile skills, and the interactive generator inside cicd-expert, remain © 2025 Viktor Farcic (MIT); see Credits.