Skip to content

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

189 Commits
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

agent-skills

CI Claude Code skills License: MIT

Personal Claude Code skills for our research group. A skill is a Markdown playbook (<skill>/SKILL.md) that Claude Code loads on demand — debugging discipline, TDD, issue triage, running ARC/T3 campaigns, Slack notifications, and more.

This repo is your skills directory: you symlink the whole thing into ~/.claude/skills, so every skill here becomes available in Claude Code everywhere.

New group member? Do the four steps below in order. Budget ~15 minutes.

Prerequisites

  • Claude Code CLI installed and signed in (claude --version works).
  • git and python3.

Setup

1. Clone and make it your skills directory

git clone https://github.com/DanaResearchGroup/agent-skills ~/Code/agent-skills
ln -s ~/Code/agent-skills ~/.claude/skills

Open a Claude Code session — the personal skills are now live. (The Slack skills need a little extra per-machine wiring; see SETUP.md.)

2. Find which skills you need to adapt

Several skills carry assumptions from the original author's machine and our group's infrastructure — hardcoded home paths, a specific Slack channel, the zeus cluster, the ARC/T3 workflow, an Obsidian vault. ADAPTATION.md is the canonical list of every such spot.

Get your personal punch-list by letting Claude Code audit the repo against it:

cd ~/Code/agent-skills
claude "Read ADAPTATION.md, then review each adaptation point against my setup. Ask me
what I actually use (Slack? Obsidian? ARC/T3 on zeus? my home directory?), then give me
an exact, file-and-line checklist of what I must change — and mark the rest N/A."

Work through the list it produces. Re-run python3 bin/lint-skills.py after edits to make sure you didn't break a skill.

3. Install gstack

gstack is a separate suite of ~23 skills (review, QA, ship, design, browse, …). Easiest path — paste this to a Claude Code session and let it run:

Install gstack: run git clone --single-branch --depth 1 https://github.com/garrytan/gstack.git ~/.claude/skills/gstack && cd ~/.claude/skills/gstack && ./setup, then add a "gstack" section to my CLAUDE.md listing the available gstack skills and telling you to use /browse for all web browsing.

Because ~/.claude/skills is this repo, gstack lands in agent-skills/gstack/ and its skills sync into top-level folders. Those are git-ignored (see .gitignore) — don't commit them; update gstack with /gstack-upgrade.

4. Install the Superpowers plugin

Superpowers is a Claude Code plugin (not a skill in this repo) that adds disciplines like brainstorming, systematic-debugging, and TDD. Install it from the official marketplace inside a Claude Code session:

/plugin install superpowers@claude-plugins-official

If that marketplace isn't registered yet, add it first, then install:

/plugin marketplace add obra/superpowers-marketplace
/plugin install superpowers@superpowers-marketplace

Verify with /plugin — you should see superpowers enabled. Full docs: https://github.com/obra/superpowers.

Optional extras

  • ui-ux-pro-max — a UI/UX design plugin (web & mobile). Only useful if you build UIs/dashboards; not needed for research work. Install with git clone https://github.com/nextlevelbuilder/ui-ux-pro-max-skill.git ~/.claude/skills/ui-ux-pro-max-skill.

What's in here

Group Skills
Notifications slack-ask, slack-notify — reach you over Slack during unattended runs
Matt Pocock (mattpocock/skills, MIT — see THIRD_PARTY_NOTICES.md) tdd, to-spec, to-tickets, implement, code-review, triage, prototype, codebase-design, domain-modeling, improve-codebase-architecture, grilling, grill-me, grill-with-docs, resolving-merge-conflicts, setup-matt-pocock-skills, writing-great-skills, handoff, edit-article, obsidian-vault, wayfinder
Workflow writing-git-commits, copilot-review — fold a PR's Copilot/security bot review back into history; drop-review — drop the review you just ran as a markdown artifact in the OS temp directory, written for the session that wrote the code; merged — after a PR lands, sync main and rebase the open PRs onto it; sync-matt-pocock-skills — pull a newer mattpocock/skills release into our vendored copies, keeping local edits
Autonomous development autodev — drive a large feature build end-to-end in one autonomous session, with adversarial Codex review at every milestone and automatic context handoff/compact/resume past 35%; spar — run an adversarial "try to break this" Codex sparring round against your current code, plan, or decision, persisted per-project; herdr — inspect and control Herdr (a terminal multiplexer for coding agents) panes, tabs, workspaces, and cross-agent communication

Plus gstack (installed in step 3) and the Superpowers plugin (step 4).

Recommended workflow

The Matt Pocock engineering skills chain into one idea → shipped flow. For a non-trivial feature, drive them in order:

  1. /grill-with-docs — interview the idea into a sharp, documented design (ADRs + glossary). For work too big to hold in one session, start with /wayfinder instead to chart it as a shared map of tickets first.
  2. /to-spec — synthesize the conversation into a spec on your issue tracker.
  3. /to-tickets — break the spec into tracer-bullet tickets, each with its blocking edges.
  4. /implement — build the tickets; drives /tdd as its red-green engine.
  5. /code-review — review the diff against your repo's standards and the spec.

Run /setup-matt-pocock-skills once per repo first, so the skills know where your issue tracker, triage labels, and domain docs live.

Other agents (Codex, Copilot CLI, Gemini)

The skills here are agent-agnostic in format: a <skill>/SKILL.md with name/description frontmatter is loaded natively by Claude Code, OpenAI Codex, Copilot CLI, and Gemini CLI alike. What changes per agent is where you symlink the repo:

Agent Skills directory (symlink target for step 1)
Claude Code ~/.claude/skills
Codex ~/.codex/skills (or the shared ~/.agents/skills)
Copilot CLI / Gemini CLI ~/.agents/skills (shared cross-runtime path)

So on Codex, do step 1 as ln -s ~/Code/agent-skills ~/.codex/skills instead. Caveats:

  • The Slack and babysit skills hard-code ~/.claude/... paths — adapt them to your agent's home (e.g. ~/.codex) per ADAPTATION.md.
  • gstack (step 3) is Claude-Code-specific — skip it on other agents.
  • Superpowers (step 4) ships its own Codex/Copilot/Gemini install — see its docs.
  • Skills that dispatch sub-agents need Codex's multi-agent tools enabled in ~/.codex/config.toml.

Maintaining this repo

  • Lint locally before pushing: python3 bin/lint-skills.py (also runs in CI on every PR).
  • CI also runs a gitleaks secret scan — never commit tokens. The Slack bot token lives outside the repo (see SETUP.md).
  • Pull vendored skill updates from upstream (Matt Pocock) with the sync-matt-pocock-skills skill — it reconciles our copies with a new mattpocock/skills release while preserving local edits.

License

MIT © 2026 Alon Grinberg Dana.

Bundled third-party skills retain their own copyright and license — see THIRD_PARTY_NOTICES.md.

Releases

Packages

Contributors

Languages