Skills

A personal collection of Claude skills — structured prompt packages that give Claude specialized capabilities for complex, multi-step tasks.

Each skill lives in its own directory and follows the Claude skill format: a SKILL.md entry point, optional agent prompts, reference documents, and helper scripts.

Available Skills

Skill	Description	Invocation
consensus-planning	Multi-agent consensus planning system.	model
design-memory	Manage a local vector store of design docs and decision records across all repos.	model
dual-adversarial-review	Run two adversarial reviewers (Claude Opus + Codex) in parallel against a spec, plan, or diff, then synthesize an accept/reject table.	model
experience-memory	Knowledge-centric self-improvement (arXiv 2607.19592): topic-keyed threads of evidence-grounded claims with stances, distilled into bundles that seed future sessions via hooks.	model
investigate	Investigate with Claude gathering evidence and you drawing the conclusions.	user
ms-frizzle	Use when the user wants to learn or understand something through hands-on experimentation and doesn't yet know what the spec or deliverable should look like — "let's poke at X", "I want to understand how Y actually works", "what would happen if", "field trip", or an explicit /ms-frizzle.	model
two-pass-review	Use before requesting human review or merging, when one review pass keeps missing a whole class of defect — deep semantic/contract bugs on one side, mechanical/local nits (null-safety, unsafe casts, unused deps, doc-vs-code drift, dead annotations, invalid generated specs) on the other.	model
wait-what	Re-pitch the last message with context, in Simplified Technical English.	user

Skill Structure

All skills live under the skills/ directory. Each skill follows this convention:

skills/
└── <skill-name>/
    ├── <skill-name>/
    │   ├── SKILL.md              # Entry point — orchestration instructions
    │   ├── agents/               # Agent prompt templates (if multi-agent)
    │   ├── references/           # Schemas, guides, templates
    │   └── scripts/              # Helper scripts
    ├── README.md                 # Human-readable documentation
    └── CLAUDE.md                 # Developer/AI-facing reference

The inner <skill-name>/ directory is what gets loaded as the skill. The outer directory holds documentation and any supporting files.

Adding a New Skill

1. Create a directory under skills/ named after the skill
2. Inside it, create another directory with the same name containing SKILL.md
3. Add agent prompts, references, and scripts as needed
4. Add a README.md alongside the inner directory documenting the skill for humans
5. Add a CLAUDE.md with architecture notes, modification guides, and testing instructions
6. Update this table with the new skill

Installing as a Plugin

This repository is a Claude Code plugin marketplace. The marketplace manifest lives at .claude-plugin/marketplace.json and exposes one plugin, patmagee-skills, which bundles every skill in the table above.

Add the marketplace and install

Run these commands inside Claude Code:

/plugin marketplace add patmagee/skills
/plugin install patmagee-skills@patmagee-marketplace

The first command registers this GitHub repository as a marketplace named patmagee-marketplace (the name comes from marketplace.json, not the repository URL). The second command installs the plugin from it. After installation, every skill is available by its name, for example /ms-frizzle or /two-pass-review.

Update

To pull the latest version of the plugin after new skills land:

/plugin marketplace update patmagee-marketplace

Disable or remove

/plugin disable patmagee-skills@patmagee-marketplace
/plugin uninstall patmagee-skills@patmagee-marketplace
/plugin marketplace remove patmagee-marketplace

disable keeps the plugin installed but inactive. uninstall removes it. Removing the marketplace also removes its plugins.

Preconfigure for a team or project

To make the marketplace and plugin available automatically for everyone working in a repository, add both keys to that repository's .claude/settings.json (or to ~/.claude/settings.json for a single user):

{
  "extraKnownMarketplaces": {
    "patmagee-marketplace": {
      "source": {
        "source": "github",
        "repo": "patmagee/skills"
      }
    }
  },
  "enabledPlugins": {
    "patmagee-skills@patmagee-marketplace": true
  }
}

Claude Code will prompt each person to trust the marketplace the first time they open the project.

Local development install

To test changes to a skill before pushing, add your local checkout as a marketplace and install from it:

/plugin marketplace add /path/to/your/checkout
/plugin install patmagee-skills@patmagee-marketplace

A local marketplace reads the working tree, so edits to a SKILL.md are picked up without a push. Remove it with /plugin marketplace remove patmagee-marketplace when you switch back to the GitHub source.

Releases

Every merge to main automatically bumps the minor version, tags the commit, and publishes a GitHub release. The bumped version is written to package.json, .claude-plugin/plugin.json, and .claude-plugin/marketplace.json together, which is what lets /plugin marketplace update detect that a new plugin version is available. Do not bump versions by hand. To land a change without cutting a release, include [skip release] in the merge commit message.

Usage

Once installed, skills trigger in two ways: Claude invokes them automatically when a task matches the skill's description, or you invoke one explicitly by typing /<skill-name>. Trigger phrases and activation conditions are documented in each skill's SKILL.md frontmatter.