Haxcms Lrs Rater

HAXcms LRS Rater

This skill runs BookLooky's official LRS content-rating rubric against a whole HAXcms site by reading its llms.txt manifest (which lists every page's .md URL in one fetch), stitching all page content into one aggregate "transcript," and then reasoning through the same rubric booklooky-lrs-rater uses for books — 10 categories (0-5), evidence-backed excerpts, and a minimum-age recommendation. No LLM API key or network call is needed for the actual rating: you (the invoking agent) are the reasoning engine, exactly as in booklooky-lrs-rater.

This skill only adds the site-specific extraction step (Step 0 below). Steps 1-7 are the same LRS methodology, scripts, and prompts as booklooky-lrs-rater — bundled here as a self-contained copy so this skill works standalone.

Why this matters (read before starting)

A HAXcms site's content is scattered across many small pages, not one linear manuscript. Rating individual pages in isolation would miss context (e.g. a "themes" page might only make sense next to the "characters" page) and would produce 10 disconnected mini-reports instead of one coherent site-level rating. So Step 0 first assembles a single ordered "transcript" from the site's outline — mirroring how a reader would actually encounter the content — and everything downstream treats that exactly like a book transcript.

Just as with booklooky-lrs-rater, every non-zero rating MUST be backed by a real quoted excerpt from the aggregated site content. reconcile-evidence.js will downgrade any unsupported positive rating to 0. Locate and quote actual page text — don't estimate.

Step 0 — Extract the site content

Two source types, both handled by the same script:

Live URL (a published HAXcms site, e.g. https://btopro.com):

node scripts/extract-site-content.js https://btopro.com <scratchDir>

Local HAXcms site checkout (a directory containing llms.txt, or site.json + pages/ as fallback):

node scripts/extract-site-content.js /path/to/site-root <scratchDir>

Either way this writes <scratchDir>/site-transcript.txt (all pages' text, each prefixed with === PAGE N: <title> ===) and <scratchDir>/page-manifest.json (index/id/title/slug/location per page, for citing which page an excerpt came from later).

How it works:

• llms.txt (preferred): The script fetches <url>/llms.txt (or reads <root>/llms.txt locally). HAXcms generates this file automatically — it lists every page's title and .md URL in its ## Pages section. One fetch gives the full page list, no site.json tree-walking or per-page HTML parsing required. The script then fetches each page's .md for clean markdown text.
• .html fallback: Some static hosts (e.g. GitHub Pages) don't serve .md files. When an .md fetch 404s, the script automatically falls back to the .html version and strips tags to plain text. This is the case for btopro.com — the script prints a note when this happens.
• site.json fallback (local only): If no llms.txt exists in a local checkout, the script falls back to reading site.json, walking the JOS outline depth-first, and reading .md (or .html) sidecars from disk.
• If a page can't be fetched at all, the script reports which page failed and continues with the rest rather than aborting the whole rating.

For very large sites, treat site-transcript.txt like a long book transcript — the chunking in Step 1 handles that the same way.

Step 1 — Chunk the transcript

node scripts/chunk-transcript.js <scratchDir>/site-transcript.txt <scratchDir>

Writes <scratchDir>/chunks.json (~100k characters per chunk, 4k overlap). Small sites will produce a single chunk — proceed the same way regardless.

Step 2 — Segment scan (per chunk, you reason directly)

For each chunk, read references/segment-scan-prompt.md and references/category-scope-rules.md, then reason through that chunk yourself to produce its signals JSON (all 10 category keys present, each an array of {quote, why}, capped at 3 per category). Write each result to <scratchDir>/scans/<chunkIndex>.json.

Do this for every chunk before moving on. If resuming across turns, skip chunks that already have a scan file.