Skills

Make a website readable by LLMs and agents at 10-50x fewer tokens by generating PageSkim .llm.md sibling files for its pages. Use when the user asks to "make my site agent-readable", "add llm.md / llms.txt-style files / PageSkim to my site", "reduce token costs for agents or AI crawlers reading my site", or "optimize my site for AI agents". Detects the site generator (plain HTML, Astro, Next.js, Hugo, Jekyll), runs pageskim generate over the build output, wires a build step plus CI validation, and emits the site index.

Make a site agent-readable with PageSkim

Goal: every published page.html gets a sibling page.llm.md (plus optional .llm.json and split files), a /.well-known/pageskim.json site index, and the build/CI wiring so siblings never go stale. Agents then read pages at 10–50x fewer tokens via the sibling files instead of raw HTML.

Everything runs through npx pageskim (published on npm) — no permanent dependency is added unless the user wants the build-step wiring.

Step 1 — detect the site generator and its output directory

Check, in order:

Signal	Generator	Build output	Build command
astro.config.*	Astro	dist/	npx astro build
next.config.*	Next.js	out/ (needs output: "export") — see note	npx next build
hugo.toml / config.toml with baseURL	Hugo	public/	hugo
_config.yml	Jekyll	_site/	bundle exec jekyll build
none of the above, .html files present	Plain HTML	the site root itself	—

Note for Next.js without static export: siblings can only cover pre-rendered pages. Offer the runtime alternative (pageskim npm package, SDK fromDocument() + expose() after hydration) for client-rendered routes, but prefer static export when the user can enable it.

Step 2 — generate siblings over the build output

bash

npx pageskim generate <output-dir> --json --split --site-index --base-url <site-url>

• Run it AFTER the site build so siblings reflect the final HTML.
• --base-url should be the production origin (used for pages without a canonical URL and for the site index).
• Exit code 2 means one or more pages look client-rendered (near-empty HTML); list them for the user and suggest pre-rendering or the SDK fallback.

Step 3 — wire it into the build

Add to package.json (or the equivalent Makefile/CI step):

json

"scripts": {
  "postbuild": "pageskim generate <output-dir> --json --split --site-index --base-url <site-url>"
}

For Hugo/Jekyll (no npm), append to the deploy script:

bash

hugo && npx --yes pageskim generate public --site-index --base-url <site-url>

Siblings are build artifacts: add them to .gitignore unless the site deploys pre-built files from git, and never hand-edit them.

Step 4 — validate in CI

Use PageSkim's composite GitHub Action, or the raw commands:

yaml

- run: npm run build
- uses: venki0552/PageSkim/.github/action@main
  with:
    path: <output-dir>
    base-url: <site-url>

Raw equivalent (also available as scripts/generate-and-validate.sh):

bash

npx pageskim generate <output-dir> --site-index --base-url <site-url>
for page in $(find <output-dir> -name '*.html' ! -name '*.llm.*'); do
  npx pageskim validate "$page" "${page%.html}.llm.md" || exit 1
done

Step 5 — verify and report

1. Pick 2–3 generated siblings and confirm: header reads sensibly, chunk anchors resolve, no nav/footer/cookie text leaked into the output.
2. Run npx pageskim validate on them (exit 0; warnings are acceptable, errors are not).
3. Report the token table the generator printed (raw HTML vs .llm.md vs header-only) so the user sees the savings — e.g. "your pages now cost agents 13x fewer tokens; the two-hop header is 150 tokens".
4. Suggest (optional): add <link rel="alternate" type="text/llm+markdown" href="page.llm.md"> to the site's head template so agents discover siblings from the HTML.

References

• Format spec: https://github.com/venki0552/PageSkim/blob/main/spec/SPEC.md
• Publisher guide (serving, caching, discovery details): https://github.com/venki0552/PageSkim/blob/main/docs/integration.md
• Live playground to demo the savings: https://page-skim.vercel.app