My personal, curated collection of agent skills. SKILL.md is an open standard, so the same folder works in Claude Code, Codex CLI, and Cursor. Some are original; many build on community prior art, credited in CREDITS.md.

Skills are grouped by domain. workflows/ holds playbooks that compose atomic skills into an end-to-end process; meta/ holds domain-agnostic skills. The domain is the install unit — install.sh installs only the domains opted into install.conf, so a non-coding domain never leaks into a coding agent. Skills can nest at any depth inside a domain (just for humans); only leaf folder names must be unique across the repo.

.
├── install.sh
├── <domain>/                 # atomic skills (e.g. software-development/)
│   └── <group>/<skill>/SKILL.md
├── workflows/                # playbooks that compose atomic skills
│   └── <workflow>/SKILL.md
└── meta/                     # domain-agnostic skills
    └── <skill>/SKILL.md

Each area has its own index:

• software-development/ — design, planning, review, engineering
• workflows/ — multi-skill playbooks
• meta/ — domain-agnostic skills

A skill's SKILL.md is the source of truth.

./install.sh            # link every configured domain into all supported agents

The standard is shared, but each agent reads a different directory — two cover all three (paths verified 2026-05-30 against the Claude Code, Codex, and Cursor docs; recheck if those tools move):

By default, install.sh links into both. Edit a skill here and every agent sees the change; once linked, each one triggers on its description: or runs as /<name>.

./install.sh --claude     # only ~/.claude/skills  (Claude Code)
./install.sh --codex      # only ~/.agents/skills  (Codex CLI)
./install.sh --cursor     # only ~/.agents/skills  (Cursor)
./install.sh --copy       # copy instead of symlink — use if an agent ignores symlinks (won't sync back)
./install.sh --force      # overwrite a foreign skill of the same name
./install.sh --uninstall  # remove only the links into this repo
./install.sh --dry-run    # preview, change nothing

Installation is default-deny — only the domains in install.conf link, so a new domain stays out of your agents until you opt it in. Each leaf SKILL.md becomes a /command named for its folder, so leaf names are unique repo-wide and the installer aborts on a collision. It won't overwrite a same-named skill it didn't create — pass --force for that.

Copy docs/skill-template.md into a kebab-case folder, add a one-line row to the area's README, and run ./scripts/validate-skills.sh until it's clean. In Claude Code, /add-skill does all of that for you.

The companion agents repo's subagent definitions reference skills from here by name (scrutinize, karpathy-guidelines, brainstorming, writing-plans, verification-before-completion, de-flaking-tests). Install skills first so those agents get the methodology they expect.

Original skills are MIT (LICENSE); adapted ones are credited in CREDITS.md.