Answer.AI Skill Plugins

** DEPRECATED** - use aai-coding repo instead.

Skills from Answer.AI for both Codex and Claude Code, built from shared sources so the two stay in sync.

Plugins

codex-aai (Codex) and claude-aai (Claude Code) carry the same skills, built from shared sources:

• persistent-python: work in a persistent clikernel IPython session - the workbench for inspection, debugging, editing, and experiments, with the pyskills tool ecosystem.
• coding-patterns: fastai coding style and conventions, including the TDD red-green testing process.
• nbdev-editing: the nbdev workflow, where notebooks in nbs/ are the source of truth.
• write-prose: write prose that doesn't read as AI slop.
• codex-docs (Codex) / claude-code-docs (Claude Code): answer questions about the tool's behavior and configuration from its official documentation; one shared source (src/skills/docs/) with per-tool names and URLs.

safecmd (Claude Code) is separate: hooks that auto-approve safe bash commands using safecmd's allowlist validation, so routine commands stop prompting for permission.

Install: Codex

Add this repository as a Codex plugin marketplace:

codex plugin marketplace add answerdotai/skill-plugins

Then open /plugins in Codex, select Answer.AI Skill Plugins, and install Codex AAI.

Plugin-installed skills are namespaced by plugin. The persistent Python skill appears as codex-aai:persistent-python.

For a plain persistent-python skill name, install the skill directly:

Use $skill-installer to install https://github.com/answerdotai/skill-plugins/tree/main/plugins/codex-aai/skills/persistent-python

Restart Codex after installing.

Install: Claude Code

Add this repository as a plugin marketplace, then install either or both plugins:

/plugin marketplace add answerdotai/skill-plugins
/plugin install claude-aai@answerdotai-skill-plugins
/plugin install safecmd@answerdotai-skill-plugins

Skills appear namespaced as claude-aai:persistent-python etc. Nothing further is needed: Claude reads a skill when its trigger applies, and safecmd's hooks run on every Bash call automatically.

Development

Each skill has ONE source, in src/skills/<name>/SKILL.md: a jinja2 template rendered once per tool with tool (codex or claude) and name (that tool's built skill name) in context. Shared text stays bare; tool-specific text goes in {% if tool == 'codex' %}...{% else %}...{% endif %} blocks or {{ }} substitutions. Tool-specific extras (e.g. agents/openai.yaml for Codex) live in src/skills/<name>/<tool>/ and are copied verbatim.

Never edit the generated files under plugins/*/skills/ - edit the source and rebuild:

./build.py          # regenerate the plugin skill trees, bumping changed plugins (--no-bump to skip)
./build.py --check  # exit nonzero if any generated file is stale (for CI)

Generated outputs are committed, since installs pull the repo as-is.

To ship a change: edit src/, run ./build.py (it bumps the version of each plugin whose output changed; --no-bump to skip), and commit. Installed plugins are cached per version, so users only see changes after a version bump plus update/reinstall. For live local use of your own checkout, skip installing the skills plugin and symlink instead, e.g. ~/.claude/skills/persistent-python -> plugins/claude-aai/skills/persistent-python: edits are then picked up on the next build, no reinstall needed.

For local marketplace testing from this checkout:

codex plugin marketplace add ~/git/skill-plugins

and in Claude Code:

/plugin marketplace add ~/git/skill-plugins

Runtime state should not be committed. Keep generated local state under ignored state/ directories.

License

Apache-2.0.