Skip to content
Conduit

skill

ai-tools

dionridley/ai-toolscommunitylistedunreviewed

Set up

Skills are installed by placing their folder where your agent loads skills.

The download comes from github.com/dionridley/ai-tools itself, not from our server, so what you get is exactly what its author published.

Install instructions come from the README

Skills do not use an MCP config. Follow the steps in the README below.

README

synced Oct 26, 2026

AI Tools

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.

Available Bundles

BundlePlugin nameDescription
Project Managementproject-managementStructured project management with research, PRDs, and implementation plans
Engineering Toolsengineering-toolsSkills and commands to assist with software engineering and coding tasks
ExperimentalexperimentalWork in Progress Tools — experimental capabilities under active development and testing

Claude Code

Via the plugin manager UI

  1. Open Claude Code and type /plugins
  2. Add the marketplace: dionridley/ai-tools
  3. Enable the plugins you want

Via the CLI

# 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-tools

Update 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 marketplace

Verify

claude plugin list

All 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.

Pi

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.

Full install

pi install git:github.com/dionridley/ai-tools

Skills 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 --all

Excluding a bundle (settings filter)

Filters 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.

Local development install

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-management

Capability notes

The 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.

Repository Layout

.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)

Contributing

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.

License

MIT

↑↓ navigate openesc close