agent-skills-bundle

agent-skills-bundle

Packaged agent skills for Claude Code and Codex. Skills are SKILL.md folders — a cross-agent standard, so the same skill works in both tools unmodified. install copies them into the directories each agent auto-discovers (.claude/skills/ and .codex/skills/), and every tool picks them up automatically.

Available Skills

Install

Two ways to install — identical result. Pick whichever fits your team.

uv (recommended — no Python required)

curl -LsSf https://astral.sh/uv/install.sh | sh   # once, if you don't have uv
uvx agent-skills-bundle install review tdd                  # named skills
uvx agent-skills-bundle install --all                       # everything
uvx agent-skills-bundle==0.2.0 install review               # pin a version

bash (local clone — no tooling required)

git clone https://github.com/priyankaiiit14/agents-skills.git
cd agents-skills
./scripts/install-local.sh /path/to/project        # omit the path to use the current dir

Both accept the --target and --scope flags described next.

Targets and scopes

--target chooses the agent(s); --scope chooses project vs machine-wide. Skills are copied to the matching directory, which each agent auto-discovers at startup:

Defaults are --target both --scope project. Use project scope to commit skills to a repo so teammates get them on git pull; use global scope to make them available in every project on your machine.

uvx agent-skills-bundle install review                    # both agents, project (default)
uvx agent-skills-bundle install review --target codex     # Codex only
uvx agent-skills-bundle install --all --scope global      # every skill, machine-wide

A lockfile records what's installed and the target/scope used — skills-lock.json (project) or ~/.agent-skills-bundle/skills-lock.json (global).

Commands

install, update, and status accept --target and --scope; update reuses the target/scope saved in the lockfile. Run any command with --help for details.

Publishing to PyPI

One-time setup

Done once, before the first release. Publishing uses PyPI Trusted Publishing (OIDC) — no API tokens are stored anywhere.

1. Create a PyPI account.
2. Confirm the project name agent-skills-bundle is free on PyPI (rename it in pyproject.toml if taken).
3. On PyPI, add a Trusted Publisher (Account settings → Publishing) with:
   • Repository: priyankaiiit14/agents-skills
   • Workflow: publish.yml
   • Environment: pypi
4. In the GitHub repo, create an environment named pypi (Settings → Environments).

Testing locally before publishing

After making changes, verify the package works before opening a PR:

Option 1 — editable install (fastest for iteration)

pip install -e .
agent-skills-bundle list
agent-skills-bundle install review --scope global

Changes to Python files take effect immediately; no reinstall needed.

Option 2 — install from a built wheel (closest to what PyPI ships)

uv build                                        # produces dist/agent_skills_bundle-<ver>-py3-none-any.whl
pip install dist/agent_skills_bundle-*.whl      # or: pip install --force-reinstall dist/...whl
agent-skills-bundle list

Use this when you want to confirm the packaged artefact (not just the source) works end-to-end.

Option 3 — run directly without installing

uvx --from . agent-skills-bundle list

Useful for a quick smoke-test in CI or a clean shell where you don't want to touch the system site-packages.

Cutting a release

Bump the version in your PR, then merge to main — publishing is automatic.

1. Bump version in pyproject.toml and src/agent_skills_bundle/__init__.py (keep them in sync).
2. Open a PR, get it reviewed, and merge to main.

The GitHub Action (.github/workflows/publish.yml) runs on every push to main. It checks whether the current package version is already on PyPI — if it is, the publish step is skipped; if it is new, the package is built and published automatically. Teams then pin with uvx agent-skills-bundle==0.2.0 install or run uvx agent-skills-bundle update to pick up the new version.

Contributing a skill

You do not need to clone the repo to use skills. Clone only to add or edit one. All skills live in src/agent_skills_bundle/skills/ — that's the single source of truth, and both agents' copies are generated from it.

git clone git@github.com:priyankaiiit14/agents-skills.git
cd agents-skills

Add a new skill

git checkout -b skills/my-skill-name
uvx agent-skills-bundle create my-skill-name   # scaffolds src/agent_skills_bundle/skills/my-skill-name/SKILL.md
# edit SKILL.md, then:
git add src/agent_skills_bundle/skills/my-skill-name
git commit -m "add my-skill-name skill"
# open a PR to main

create writes the folder with the required frontmatter already filled in. Then add the skill to the table at the top of this README.

Update an existing skill

git checkout -b skills/my-skill-name        # or fix/my-skill-name for a bug fix
# edit src/agent_skills_bundle/skills/my-skill-name/SKILL.md (or its references/, scripts/, assets/)
git add src/agent_skills_bundle/skills/my-skill-name
git commit -m "update my-skill-name: <what changed>"
# open a PR to main

The change ships to users in the next release (see How releases work). Users pull it with uvx agent-skills-bundle update.

Branch naming

Skill layout

src/agent_skills_bundle/skills/
  <skill-name>/
    SKILL.md          ← required; keep concise
    references/       ← large or conditional reference material
    scripts/          ← executable helpers
    assets/           ← templates, binaries, examples

Every SKILL.md needs YAML frontmatter — name and description are required (the description is what shows in agent-skills-bundle list and what each agent uses to decide when to trigger the skill):

---
name: my-skill-name
description: One-line description of what this skill does and when to invoke it.
---