Docs Skills

docs-detect-source — Identify and route a docs source

Workflow#

  1. Inspect the input: classify as a plain URL, a GitHub repo URL, or a local path.
  2. For a plain URL — fetch the page and look for platform signals in HTML meta tags, CDN links, and domain name. If no platform matches, classify as a generic website.
  3. For a GitHub repo URL — fetch root contents via the GitHub API and look for known config files (mint.json, SUMMARY.md, docusaurus.config.js, .vitepress/, astro.config.*, etc.). If only a README and source directory are present, classify as a code repo.
  4. For a local path — inspect the directory structure for the same config file signals.
  5. Map the detected type to a downstream skill and report both.

Guardrails#

  • For detection (platform signals in meta tags / CDN links), a plain HTTP fetch of the raw HTML is enough — those signals live in the shell. But note: a non-empty HTML shell does not mean the content is there. Most product sites are JS SPAs whose <main> is empty until rendered, so the downstream builder (/docs-from-site) must render with a browser to read real content — do not let a 200 with populated <head> fool the content pass into skipping the render.
  • A single ambiguous signal is not enough — require at least one platform-specific config file or meta tag before assigning a platform type.
  • If detection is inconclusive, default to website / /docs-from-site rather than blocking.
  • Never mutate the source; detection is read-only.

Skill Routing#

Detected type Downstream skill Subagent
website /docs-from-site docs-site-crawler (Haiku)
github-code-repo /docs-from-code docs-code-crawler (Haiku)
Any recognized docs platform /docs-from-docs docs-platform-importer (Haiku)

Acceptance Criteria#

  • Input classified into exactly one type without prompting the user
  • Platform identified from signals (meta tags, config files, domain) rather than guessing
  • Downstream skill name reported alongside the detected type
  • Inconclusive inputs default to website gracefully