Przejdź do treści
Biblioteka skilli
Programowanie · Konwersja i formaty

docs-crawler

/docs-crawler

Zbiera całą stronę dokumentacji lub design systemu w jeden korpus Markdown dla LLM: odkrywa strony z sitemap.xml, wyciąga treść, pobiera obrazy i renderuje strony z JavaScriptem.

Repozytorium
ko-design-md ↗
Licencja
MIT
Ostatnia zmiana repozytorium
25.10.2026

Ocena i bezpieczeństwo

87/ 100Klasa: dobry
Format20 / 25
Frontmatter z nazwą i opisem, poprawna nazwa, opis do 1024 znaków, nazwa zgodna z folderem.
Odkrywalność17 / 25
Opis mówi, kiedy użyć skilla, ma sensowną długość i jest w 3. osobie.
Treść35 / 35
Co najmniej 150 słów i do 500 linii, nagłówki, listy kroków, przykłady albo kod.
Higiena15 / 15
Bez zaślepek typu TODO, bez ścieżek Windows, z licencją.
Ryzykobrak sygnałów w skanie
Pliki w paczce
17
Skrypty
14
Komentarze HTML
0
Adresy URL w SKILL.md
1
Przestarzałość

Brak sygnałów. Repozytorium zmienione 25.10.2026, bez odwołań do wycofanych modeli i API.

Licencja

MIT, otwarta: można używać i zmieniać, z zachowaniem informacji o autorze.

Co sprawdzić przed instalacją
  • 14 skryptów w paczce. Przejrzyj je przed instalacją, bo Claude może je uruchomić.

Skrypty: scripts/crawl.ts, scripts/crawl/args.test.ts, scripts/crawl/args.ts, scripts/crawl/corpus.test.ts, scripts/crawl/corpus.ts, scripts/crawl/discover.test.ts, scripts/crawl/discover.ts, scripts/crawl/extract.test.ts

Co robi i kiedy

Zbiera całą stronę dokumentacji lub design systemu w jeden korpus Markdown dla LLM: odkrywa strony z sitemap.xml, wyciąga treść, pobiera obrazy i renderuje strony z JavaScriptem.

Wyzwalacze z opisu

s main content as clean Markdown, downloads each page
Oryginalny opis (EN)

Crawl an entire documentation or design-system website into one LLM-ready Markdown corpus. Discovers every page from the site's sitemap.xml (with a same-origin link-following fallback when there is no sitemap), extracts each page's main content as clean Markdown, downloads each page's images (and inline base64 images) into a local folder referenced by relative paths, and renders JavaScript-heavy pages with a headless browser. Use this whenever someone wants the WHOLE of a multi-page docs site, design-system site, API reference, component library, or knowledge base turned into Markdown — for ex

Jak działa

  1. When to use
  2. Inputs
  3. Running the crawl
  4. Output
  5. After crawling
  6. Output location and the design-md skill
  7. Notes and edge cases
Sekcje z nagłówków SKILL.md, w oryginale · 112 linii, 726 słów
Podgląd SKILL.md
---
name: docs-crawler
metadata:
  internal: true
description: >-
  Crawl an entire documentation or design-system website into one LLM-ready
  Markdown corpus. Discovers every page from the site's sitemap.xml (with a
  same-origin link-following fallback when there is no sitemap), extracts each
  page's main content as clean Markdown, downloads each page's images (and
  inline base64 images) into a local folder referenced by relative paths, and
  renders JavaScript-heavy pages with a headless browser.
  Use this whenever someone wants the WHOLE of a multi-page docs site,
  design-system site, API reference, component library, or knowledge base
  turned into Markdown — for example "crawl this docs site", "turn this design
  system site into markdown", "make an LLM or RAG corpus from these docs",
  "mirror the entire documentation", "archive the whole knowledge base", or
  "give an AI the full X documentation". It applies even when the user never
  says the word "crawl" — any request to capture, ingest, archive, or convert
  a whole site's documentation pages into text fits. The design-md skill also
  calls it during research to gather a brand's published design-system docs.
  Do not use it for reading or summarizing a single page, for scraping
  structured data such as prices into a spreadsheet, or for generating a
  sitemap.
---

# docs-crawler

Crawl a documentation website into a single Markdown corpus that an LLM can
consume as context.

## When to use

The user wants to capture a multi-page documentation, design-system, or
reference site as Markdown — to feed an LLM, archive it, or research a design
system. This skill is also invoked by the `design-md` skill's research phase.

## Inputs

- **Site URL** (required) — the documentation site, e.g.
  `https://socarframe.socar.kr/`. Any page on the site works; page discovery
  starts from the site's `sitemap.xml`.
- **Output directory** (optional) — where to write the corpus. Defaults to
  `./{host}-docs/` in the current working directory.

If the user has not provided a URL, ask for one — do not guess.

## Running the crawl

From the repository root:

```
pnpm crawl:docs <site-url> [--out <dir>] [--external-images]
```

Equivalent direct form:

```
pnpm exec tsx .claude/skills/docs-crawler/scripts/crawl.ts <site-url> [--out <dir>]
```

The engine discovers pages via `sitemap.xml` (falling back to same-origin
link-following), fetches each page, extracts the main content, and converts it
to Markdown. It prints per-page progress and a final summary line.

By default it also **localizes images**: every external image and inline base64
`data:` image is downloaded into `crawl/images/` and the Markdown is rewritten
to relative paths, so the corpus is self-contained for renderers that can't
fetch external URLs (Claude Design, offline previews). Pass `--external-images`
to skip downloading and keep the original external URLs instead.

## Output

These artifacts are written under the output directory:

- **`crawl-corpus.md`** — every successful page merged into one document, with
  a table of contents and a `Source:` URL per page. This is the primary
  deliverable.
- **`crawl/pages/{NNN}-{slug}.md`** — one file per page, each with
  `source_url` / `title` / `method` frontmatter.
- **`crawl/images/`** — downloaded image files (skipped with
  `--external-images`), referenced from the corpus and page files by relative
  paths.
- **`crawl/manifest.json`** — a per-URL audit log (status, fetch method,
  extracted character count, error reason).

## After crawling

Report back to the user:

- How many pages succeeded versus failed (from the summary line or
  `crawl/manifest.json`).
- The path to `crawl-corpus.md`.
- If any pages failed, list them from the manifest with the reason (e.g.
  `disallowed by robots.txt`, `no extractable content`).

Do not paste the whole corpus into the conversation — point the user at the
file. The corpus can be large.

## Output location and the design-md skill

The `--out` argument controls where everything is written. When the `design-md`
skill calls this crawler during its research phase, it passes `--out` pointing
at its cache directory (`.claude/cache/design-md/{slug}/`), so the resulting
`crawl-corpus.md` (and the localized `crawl/images/`) land where
`research-collector` can read and cite them. The whole cache directory is
gitignored, so downloaded images stay out of version control. Nothing special
is needed for that case — just run the crawl with the given `--out`.

## Notes and edge cases

- **No sitemap** — discovery automatically falls back to following same-origin
  links from the start URL.
- **JavaScript-rendered sites** — if a page returns an empty shell, the engine
  re-fetches it through a headless browser. Chromium is installed automatically
  on first need (~150 MB, one-time). Static sites never launch a browser.
- **Images** are downloaded into `crawl/images/` by default and referenced by
  relative paths, so the corpus is self-contained — this covers external
  `<img>` URLs and inline base64 `data:` images alike. A non-base64
  (percent-encoded) `data:` URI, or any image whose bytes fail to download or
  decode, instead becomes an `inline-image-omitted` placeholder (alt and title
  text are kept) so raw base64 never bloats the corpus. With `--external-images`
  the crawler downloads nothing: external URLs stay as-is and inline `data:`
  images collapse to the placeholder.
- **Few or no images is normal, especially for design-system sites** — modern
  docs and design-system sites (Docusaurus and similar) render icons and
  component visuals as inline `<svg>`, CSS, or live DOM rather than `<img>`
  files; any real `<img>` (e.g. a navbar logo) is usually site chrome the
  content extractor correctly drops. An image-light or image-free corpus is
  expected in that case — it is not a crawl failure. Capture visuals separately
  (screenshots) if the research needs them.
- **robots.txt** is respected; disallowed paths are skipped.
- **Page cap** — at most 200 pages per crawl.
- **Zero pages crawled** — the engine exits non-zero and prints the reason
  (bad URL, blocked, or an empty site). Surface that to the user rather than
  reporting success.
Pliki w paczce (17)
  • SKILL.md 6 KB
  • trigger-eval-queries.json 3 KB
  • evals/evals.json 4 KB
  • scripts/crawl.ts 11 KB
  • scripts/crawl/args.test.ts 2 KB
  • scripts/crawl/args.ts 3 KB
  • scripts/crawl/corpus.test.ts 3 KB
  • scripts/crawl/corpus.ts 3 KB
  • scripts/crawl/discover.test.ts 3 KB
  • scripts/crawl/discover.ts 7 KB
  • scripts/crawl/extract.test.ts 7 KB
  • scripts/crawl/extract.ts 5 KB
  • scripts/crawl/fetch-page.test.ts 1 KB
  • scripts/crawl/fetch-page.ts 5 KB
  • scripts/crawl/images.test.ts 8 KB
  • scripts/crawl/images.ts 10 KB
  • scripts/crawl/types.ts 2 KB

Wersje

W bazie nie ma innych skilli o nazwie docs-crawler.

Migawka bazy z 13.11.2026 · wersja paczki 26341b13eb3d (skrót SHA-256 całej paczki)

Pobrania w ostatnich 30 dniach: 0 · ranking

Zgłoś problem z tym skillem: szkodliwa treść, błędny opis albo prośba autora o usunięcie.

Podobne skille

  • 100scrapling
    Pobiera strony i wyciąga z nich ustrukturyzowane dane w Pythonie, od pojedynczego żądania HTTP po pełny crawl, także ze stron renderowanych JavaScriptem i z selektorami odpornymi na redesign.
  • 100algolia-crawler
    Indeksuje strony lub całe witryny w Algolia przez CLI i Crawler, na przykład pod RAG, wyszukiwanie semantyczne i Agent Studio, z pisaniem i debugowaniem konfiguracji crawlera.
  • 100webforai
    Zamienia strony WWW albo lokalne pliki HTML na czysty Markdown gotowy dla LLM, także strony z JavaScriptem przez hostowaną platformę.
  • 100read-web-page
    Pobiera publiczną stronę jako czysty markdown przez przeglądarkę SnapRender, także gdy treść powstaje w JavaScript.
  • 95docs-page-generator
    Pomaga zaplanować i zoptymalizować stronę dokumentacji: strukturę, sekcje Getting Started, API Reference i poradniki.
  • 100read-web-page
    Pobiera publiczną stronę jako czysty markdown przez przeglądarkę SnapRender, przydatne przy dokumentacji i changelogach.

Uwagi i zgłoszenia

Wiadomość trafi do Bartosza Sokolińskiego. Adres e-mail jest potrzebny tylko do odpowiedzi. Szczegóły w polityce prywatności.