Cross-harness AI agent tooling: one repo that serves as both a Claude Code plugin marketplace and a Pi package, built from shared, harness-neutral skill bundles. No skill file exists twice — each bundle under bundles/ carries both harness manifests (.claude-plugin/plugin.json for Claude Code, package.json for Pi) over the same skills.
This repo supersedes
dionridley/claude-plugins.
| Bundle | Plugin name | Description |
|---|---|---|
| Project Management | project-management |
Structured project management with research, PRDs, and implementation plans |
| Engineering Tools | engineering-tools |
Skills and commands to assist with software engineering and coding tasks |
| Experimental | experimental |
Work in Progress Tools — experimental capabilities under active development and testing |
- Open Claude Code and type
/plugins - Add the marketplace:
dionridley/ai-tools - Enable the plugins you want
# Add the marketplace (URL, path, or GitHub repo)
claude plugin marketplace add dionridley/ai-tools
# Install plugins — all optional, pick what you use
claude plugin install project-management@ai-tools
claude plugin install engineering-tools@ai-tools
claude plugin install experimental@ai-toolsUpdate and remove:
claude plugin marketplace update ai-tools # refresh the catalog
claude plugin update project-management # update one plugin
claude plugin uninstall project-management # remove one plugin
claude plugin marketplace remove ai-tools # remove the marketplaceclaude plugin listAll three plugins should show as enabled. The skills then appear inside a Claude Code session as slash commands — the project-management plugin provides /dr-init, /dr-research, /dr-prd, /dr-plan, and /dr-ship; the experimental plugin provides /mvp.
The root package.json declares this repo as a Pi package — its pi.skills manifest globs bundles/*/skills/*, so one install brings in every bundle's skills.
pi install git:github.com/dionridley/ai-toolsSkills auto-discover by description; explicit invocation is /skill:<name> (e.g. /skill:dr-plan). Skills marked disable-model-invocation: true (dr-init, dr-research, dr-ship, mvp) load but are explicit-only — hidden from auto-discovery by design.
Update pi itself and all installed packages:
pi update --allFilters layer on top of the manifest — they narrow what the install declared. To take everything except a bundle, use the object form of the packages entry in ~/.pi/agent/settings.json (global) or .pi/settings.json (per project):
{
"packages": [
{
"source": "git:github.com/dionridley/ai-tools",
"skills": ["!bundles/experimental/**"]
}
]
}Filter syntax: globs select, !pattern excludes, [] loads none of that resource type, an omitted key loads all. pi config edits the same settings interactively (Tab switches global ↔ project). Bundle boundaries are directory boundaries, so filters read like plugin names.
A directory install is a live reference, not a copy — working-tree edits are visible in the next Pi session:
# whole repo
pi install /path/to/ai-tools
# a single bundle (per-bundle manifests activate on direct installs)
pi install /path/to/ai-tools/bundles/project-managementThe skills are written capability-conditionally: Claude Code features are used when present, with documented degradations elsewhere. On Pi (and other Agent Skills harnesses), expect:
- dr-research — requires web search/fetch; install a web-access package before using it.
- dr-ship — requires git; the PR step needs the GitHub CLI (
gh) and a GitHub remote, otherwise it displays the PR content for manual use. - mvp — runs in Reduced Sequential Mode (tasks execute inline and sequentially instead of via parallel subagents/worktrees) with single-session setup; requires git.
- Everything else (dr-init, dr-plan, dr-prd, frontend-design, react-19) — runs on stock capabilities; structured-question and verifier-subagent steps fall back to plain text and inline self-review.
.claude-plugin/marketplace.json # Claude catalog — sources point at ./bundles/*
package.json # Pi catalog (private) — pi.skills globs bundles/*/skills/*
bundles/<name>/ # one bundle = one plugin: skills/, agents/, both manifests
pi/ · claude/ # cross-cutting harness-exclusive artifacts (.gitkeep until first content)
See AGENTS.md for development guidelines, the repository structure, and the release ritual — it is the canonical instruction file for every coding agent working here (CLAUDE.md is a thin pointer to it).
MIT