ai-hats-library

ai-hats-library

The agent library for ai-hats — the portable content the framework composes into role prompts: skills, roles, traits, and rules, plus the runtime hooks they ship. It is a data-only package (YAML + Markdown + a few hook scripts); it holds no framework runtime code and has zero dependencies.

Two ways to use it

1. Drop-in for any Claude Code user (no ai-hats, no pip). The skills are in the standard Claude Code Agent-Skill format (SKILL.md frontmatter). Clone this package and copy the ones you want into ~/.claude/skills/:

git clone https://github.com/muratovv/ai-hats
cp -r ai-hats/packages/ai-hats-library/src/ai_hats_library/core/skills/systematic-debugging ~/.claude/skills/

The ai-hats composition metadata rides in a namespaced ai_hats: frontmatter key that non-ai-hats agents simply ignore — so every SKILL.md stays drop-in.

2. As the content dependency of ai-hats. ai-hats pins this package (ai-hats-library>=…) and resolves its built-in library layer from it via importlib.resources. The precedence stack layers user overrides on top, last-wins: built-in (this package) < ~/.ai-hats < project libraries.

Layout

src/ai_hats_library/
  core/     # engine fundament — universal skills, base traits, global rules, pipelines
  usage/    # opinionated catalog — domain roles, dev::/env:: traits, opt-in skills
  hooks/    # runtime hook sources materialized into consuming projects

core/ and usage/ are the two library layers (core first = lowest priority). Local development against ai-hats uses the AI_HATS_LIBRARY_ROOT environment override pointed at a checkout of this package.

Versioning

Semantic versioning on the format schema (skill frontmatter + composition schema + resolver expectations), decoupled from ai-hats's own version. The shipped schema_version (in manifest.yaml) is the contract; ai-hats pins a compatible range and fails loud on a library newer than it understands. MAJOR = breaking schema change, MINOR = skills/roles added, PATCH = content fix.

License

MIT — see LICENSE. Third-party-derived skills carry their own upstream attribution in-tree (metadata.yaml upstream: blocks + per-skill LICENSE files); see the repository-root THIRD_PARTY_NOTICES.