Agent skills in the cross-agent SKILL.md format. Use them to operate a product community, explain a pull request, or write grounded technical documentation.
Any agent, via the skills CLI (browse the ecosystem at skills.sh):
npx skills add naxodev/skillsAdd -a <agent> to target a specific agent, e.g. -a opencode or -a claude-code.
Claude Code, as a plugin:
/plugin marketplace add naxodev/skills
/plugin install skills@naxodev
| Skill | Use it for | Output | Invocation |
|---|---|---|---|
discord-community-server |
Plan, audit, improve, or launch a product community | Operating plan matched to member journeys and staff capacity, approved configuration, or read-only audit with evidence and limits | Automatic when relevant |
pr-walkthrough |
Explain the motivation and mechanics of a PR | Self-contained HTML with a consistent four-part arc and final recap, with optional audio | Explicit request |
writing-technical-docs |
Write or improve how-to guides, tutorials, reference, and explanations | Reader-focused documentation in one quadrant, grounded in evidence and verified against its promised outcome | Automatic when relevant |
Automatic invocation depends on the host agent. The walkthrough disables implicit invocation in its Claude and Codex metadata.
- Discord: “Audit our beta community's roles and onboarding. Report gaps without changing the server.”
- PR walkthrough: “Use pr-walkthrough for PR #42. Make it text-only and explain the trade-offs.”
- Technical docs: “Write a how-to for this package's public API. Read the exports and verify the example.”
See the sample walkthrough, including downloadable HTML and its editable manifest. It uses a fictional PR and requires no GitHub account to view.
- Discord: repository and product context; official documentation access for current capabilities. Live configuration needs a human-authenticated browser session, approved changes, and human-operated member checks. Offline audits report unavailable checks as untested.
- PR walkthrough: Node.js 22+, authenticated GitHub CLI (
gh), and an agent with subagent support. Install optional rendering and audio dependencies withbun installinskills/pr-walkthrough/scripts/. Audio also needsffmpegwithlibmp3lame; the first audio run downloads the model. Missing optional dependencies produce text or plain-code fallbacks. - Technical docs: access to relevant source, configuration, or authoritative external documentation. Verification uses the project's available checks and reports any blocked live-service checks.
Use Node.js 22+ and Bun 1.3.7. Install the contributor dependencies from the repository root:
bun install --frozen-lockfile
bun install --frozen-lockfile --cwd skills/pr-walkthrough/scripts
npm install --global skills-ref@0.1.5 @anthropic-ai/claude-code@2.1.233The root dependencies provide linting, type checking, and catalog checks. The skill has its own lockfile so installed copies can set up their bundled scripts independently.
Run all required checks:
bun run checkThis runs strict JavaScript/TypeScript checking, lint, deterministic tests, catalog and local Markdown link checks, skill metadata validation, and strict plugin validation.
To run metadata and catalog validation alone, use bash scripts/validate.sh.
Run bun run sample to rebuild the sample HTML. No audio synthesis is needed.
For instruction changes, follow the behavior evaluation guide. Those evaluations require fresh agent sessions and evidence-based review; CI does not claim a model behavior pass rate.
The experimental narrative audit command prepares text-field coverage and applies source-linked audit edits to a separate manifest. It is not part of the default walkthrough workflow.
Contributors should follow the repository contract in AGENTS.md. See conceptual improvement ideas for proposed experiments and ways to measure their value.