Scrap

scrap

Scrapes the public catalog of any online store into a PDF + HTML report, and compares scraps from different dates to estimate sales, revenue and trend. The platform (Tienda Nube, Shopify, WooCommerce, or generic structured-data stores) is detected automatically — the user never has to say which one it is.

{SKILL_DIR} = this skill's directory (passed when invoked, e.g. ~/.claude/skills/scrap). Run the commands from the user's current directory (not the skill's) so output lands in ./output near them. Reports are generated in Spanish (es-AR formatting).

Setup (once)

If {SKILL_DIR}/node_modules/puppeteer-core doesn't exist, install dependencies:

npm install --prefix "{SKILL_DIR}" --no-audit --no-fund

Requires Node ≥18 and Google Chrome/Chromium (auto-detected on Mac/Windows/Linux; override with CHROME_PATH). Chrome is only needed for the PDF — --html-only works without it.

Choosing the mode

The first argument decides:

• starts with http/a domain → catalog mode (scrape that store).
• is compare → comparison mode (uses the attached reports).

Catalog mode — /scrap <url>

node "{SKILL_DIR}/scripts/scrape.js" "<url>"

Generates ./output/Scrap_<Brand>_<date_time>.pdf + .html. Report the real output to the user: both absolute paths + product/stock summary + which platform was detected.

Notes by platform (the script handles this alone; just relay it):

• Tienda Nube: full data — numeric stock per variant and sold_qty (units sold).
• Shopify / WooCommerce / generic: prices and availability, but no numeric stock (those platforms don't make it public). Sales comparison will be availability-based.
• If the store can't be detected or has no structured data, the script says so clearly.

Useful flags: --limit N (quick test), --fresh (ignore cache), --html-only (skip PDF), --json (also export the raw dataset), --lang es|en (report language, default es — only translates interface labels, never product names/descriptions), --out DIR. Without Chrome installed the script still delivers the HTML (it warns and skips the PDF). If the user asks for the report in English (or their message is in English), pass --lang en.

Comparison mode — /scrap compare

The user attaches 2+ reports (PDF or HTML) generated earlier with /scrap, from the same store.

1. Prepare each attached file:

• If it's an HTML from this skill: nothing to do — compare.js accepts catalog HTML files directly and extracts the embedded dataset itself.
• If it's a PDF from this skill: build a JSON dataset by hand: read the cover ("Fecha y hora del scrap" → scrapedAt) and the "Detalle por producto" section. Per product build { "id", "name", "price" (number), "stock" (total or null), "soldQty" ("Vendidas" row if present), "available" }. Save it to ./output/.compare/<n>.json.
• If it's an old/external PDF without a date: ask the user for that scrap's date (and time).

2. Run the comparison — HTML files and JSON datasets can be mixed freely, and can come from scraps in either language (the report language is independent from the input):

node "{SKILL_DIR}/scripts/compare.js" reporte1.html ./output/.compare/2.json [...] --lang en

Generates ./output/Reporte_Ventas_<Brand>_<date_time>.pdf + .html. Give the user the paths + the headline (estimated units and revenue — or, for stores without numeric stock, which products sold out in the period).

Rules

• Always deliver both PDF and HTML and report absolute paths. The HTML is the primary deliverable (lightweight, full data, machine-readable dataset embedded); highlight it.
• Light PDF by default: render.js downsizes images to 320px for the PDF (the HTML keeps full quality) so big catalogs don't produce 100+ MB PDFs. Don't raise PDF_IMG_SIZE unless explicitly asked.
• The comparison is an estimate from public data (not the store's real books): with numeric stock it's a floor; without it, it only detects sold-out transitions. Say so when delivering, as the report itself does.
• More scraps, closer together = better estimates. Suggest scraping every few days.
• Don't invent data: if a PDF has no date, ask; if a store exposes nothing, say so.

Technical detail

• Platform detection: scripts/lib/detect.js. Adapters: scripts/lib/platforms/*.js (one per platform, same normalized product shape). Sales math: scripts/compare.js.
• Local page cache in ./output/.cache/<host>/ — re-runs are fast and a closed Tienda Nube store (password page) can regenerate from cache. --fresh forces re-download.
• Tests: npm test (fixtures with real store data live in test/fixtures/).