Agent skills

leadtype emits a discoverable SKILL.md surface — the open Agent Skills format (originally Anthropic's, supported by Claude Code, Cursor, Codex, Copilot, Gemini CLI, and others). A skill is a name, a description, and instructions an agent loads on demand; the surface lets a client discover yours without installing anything.

It's on by default: every leadtype generate emits an auto docs-skill that points an agent at your docs. Declare more in docs.config.ts when you have real capability skills to expose.

What gets emitted

In site mode (default), under your output root:

• /.well-known/agent-skills/index.json — the discovery manifest: each skill's name, description, path, and a sha256-… integrity hash. Clients read this first (progressive disclosure — name + description before the full instructions).
• /.well-known/agent-skills/<name>/SKILL.md — one per skill (agentskills.io frontmatter + Markdown instructions).
• /.well-known/agent-card.json — an A2A agent card: name, description, url (your MCP endpoint when enabled, else the site), version, capabilities, defaultInputModes/defaultOutputModes, and each skill as { id, name, description, tags }. Its provider reuses your top-level organization and documentationUrl is product.docs (default <baseUrl>/docs) — so leadtype's own card reads provider: { organization: "Inth", url: "https://inth.com" }, documentationUrl: "https://leadtype.dev/docs".

In bundle mode (--bundle), a single SKILL.md ships at the package root next to AGENTS.md, so an agent reading node_modules/<pkg> finds it on the filesystem the same way it finds AGENTS.md. That single file is always the auto docs-skill (the offline docs pointer); capability items are a site-only surface and aren't bundled. With docsSkill: false there's nothing to point at, so bundle mode emits no SKILL.md.

These are static files. leadtype never hosts a skills endpoint or an A2A runtime — your site serves the files, the same as llms.txt.

What is the auto docs-skill?

A <product>-docs skill whose instructions are a pointer to your docs, adapted to whichever surface exists:

• Bundle: read ./AGENTS.md → the per-topic Markdown in ./docs/ (offline, version-matched).
• Site: read /llms.txt, and — when agents.mcp.enabled is set — connect the docs MCP server for targeted retrieval.

It stays a thin pointer, not a third copy of the docs; progressive disclosure means an agent loads the body only when a task matches. It describes reading your docs — not driving your product's API. Capability skills are yours to declare.

How do I declare my own skills?

Add agents.skills.items in docs.config.ts. Each becomes its own SKILL.md:

import { defineDocsConfig } from "leadtype";

export default defineDocsConfig({
  product: { name: "Acme", tagline: "…" },
  agents: {
    skills: {
      items: [
        {
          name: "deploy-acme",
          description: "Deploy an Acme app to production. Use when the user asks to ship or release.",
          license: "MIT",
          allowedTools: ["Bash", "Read"],
          bodyPath: "skills/deploy.md", // or `body: "...inline markdown..."`
        },
      ],
    },
  },
});

bodyPath is resolved against your docs source root (the --src you pass generate). Keep skill bodies outside your linted docs content tree — they're build inputs, not pages, so leadtype lint would otherwise flag them for missing page frontmatter. A sibling skills/ directory works well. Set docsSkill: false to drop the auto docs-skill (e.g. you wrote a richer one yourself), or agents.agentCard.enabled: false to skip the A2A card.

Config

agents: {
  skills: {
    docsSkill: true,   // auto "use these docs" skill (default true)
    items: [ /* DocsSkillSpec[] */ ],
  },
  agentCard: { enabled: true }, // emit /.well-known/agent-card.json (default true)
  mcp: { enabled: true },       // makes the site docs-skill point at your MCP endpoint
}

The agent card's provider comes from the top-level organization and its documentationUrl from product.docs — see product (identity).

A SKILL.md's frontmatter follows agentskills.io: name and description are required; license, compatibility, allowed-tools, and metadata are optional.