agent-config

Portable skills and agents for AI coding agents.

Install

git clone https://github.com/unvalley/agent-config.git
cd agent-config
just install

just install links this repository's assets, restores third-party skills from the skills.sh lock, and installs the git hooks. On a new machine run chezmoi init --apply unvalley first, so the dotfiles and that lock are in place. just --list has the rest: status, uninstall, third-party, validate, eval, and eval-all.

Assets are symlinked, so local edits apply immediately. Claude Code receives skills/ and agents/; Codex receives skills/ through ~/.agents/skills. The installer is a Rust CLI behind those recipes; cargo run -- install --help covers its --copy, --force, and --dry-run flags.

Codex configuration (~/.codex/config.toml) and the global skills lock (~/.agents/.skill-lock.json) belong in dotfiles, not here.

Skills

Third-party skills are managed by skills.sh and restored by just third-party. Run chezmoi add ~/.agents/.skill-lock.json after changing them so the lock stays in dotfiles.

This repository's own skills can also be installed elsewhere:

npx skills add unvalley/agent-config
gh skill install unvalley/agent-config/skills/design-principles

Authoring a skill

1. Create skills/<name>/SKILL.md with name and description frontmatter.
2. Add agents/openai.yaml with display_name, short_description, and a default_prompt that invokes $<name>.
3. Put optional detail in references/, deterministic code in scripts/, and output resources in assets/.
4. Keep frontmatter ASCII-only and validate with just validate <name>.
5. Add eval cases under evals/<name>/, as described below.

Evals

Skills are evaluated with Claude Code's claude plugin eval. A case lives in evals/<skill>/<case>/:

• prompt.md holds the prompt; its plugins frontmatter names the skills to load, so a routing case can load a sibling skill too.
• Each file in graders/ is one check: regex on the reply, tool_used for whether a skill fired, or llm for a rubric.

just eval git-commits   # one skill
just eval-all           # every skill

Each case runs with and without its skills, so the report shows what the skill adds. Flags pass through to claude plugin eval, such as --runs 1 or --tag routing. Runs use your Claude Code login, and reports stay local in evals/results/. The latest scores are in skills/benchmark.md.