Deep-crawl any website from start URLs, return per-page LLM-ready text/markdown/HTML plus metadata (title, description, author, language, canonical URL, OG) and in-scope outbound links. Use when user mentions deep crawl website, recursive crawl, crawl a whole site, scrape entire website, scrape docs site, scrape documentation, scrape knowledge base, scrape blog, build RAG corpus, build vector database from website, knowledge base for chatbot, GPT knowledge files, llms.txt, sitemap crawl, BFS crawl, scrape with depth or page limit, include exclude URL globs, remove boilerplate, strip navigation header footer, website to markdown, website to text, multi-page extraction, bulk page scraping, clean markdown from URL, docs site to markdown corpus, site to clean corpus. Also applies to building RAG pipelines, indexing a customer site, syncing docs into a vector store, generating training corpora from any docs hub, or expanding a single start URL into a clean corpus of every reachable in-scope page.
Deep-crawl any website from start URLs, return per-page LLM-ready text/markdown/HTML plus metadata (title, description, author, language, canonical URL, OG) and in-scope outbound links. Use when user mentions deep crawl website, recursive crawl, crawl a whole site, scrape entire website, scrape docs site, scrape documentation, scrape knowledge base, scrape blog, build RAG corpus, build vector database from website, knowledge base for chatbot, GPT knowledge files, llms.txt, sitemap crawl, BFS crawl, scrape with depth or page limit, include exclude URL globs, remove boilerplate, strip navigation header footer, website to markdown, website to text, multi-page extraction, bulk page scraping, clean markdown from URL, docs site to markdown corpus, site to clean corpus. Also applies to building RAG pipelines, indexing a customer site, syncing docs into a vector store, generating training corpora from any docs hub, or expanding a single start URL into a clean corpus of every reachable in-scope page.
Website Deep Crawl
Input: one or more start URLs (+ optional scope, depth, page-count, globs, removal selectors). Output: per-page records {url, crawl, metadata, text, markdown, html, outboundLinks} for every page reached within scope.
Language
All process output to user (progress updates, process notifications) follows the user's language.
Objective
From a small set of start URLs, breadth-first crawl every reachable in-scope page, strip boilerplate (navigation, header, footer, cookie banners, etc.), and emit per-page LLM-ready content (text / markdown / HTML) plus structured metadata — suitable for feeding RAG pipelines, vector databases, or chatbot knowledge bases.
Prerequisites
One or more start URLs are provided by the caller.
Target pages are publicly reachable, OR the running browser is already logged in for any pages behind authentication.
A working directory is available for writing per-page JSON records and the crawl state file.
Pre-execution Checks
1. Tool Readiness
If browser-act has been confirmed available in the current session → skip this step.
Invoke browser-act via Skill tool to load usage. If installation or configuration issues arise, follow its guidance to resolve then retry.
2. Login Verification (when prerequisites include login requirement)
If login status for the target site has been confirmed in the current session → skip this step.
Otherwise: open the target site and observe the page login status:
Logout/sign-out entry, user avatar, or username exists → logged in, continue execution
Login/register entry exists with no logout entry → not logged in, inform the user that login is needed first, assist the user in completing the login flow
User refuses or cannot log in → terminate execution.
Capability Components
This Skill's operational boundary = what the user can manually do in their browser. It only reads data already displayed to the user on the page, never bypassing authentication or access controls. Its role is equivalent to copy-pasting on the user's behalf — the data is already on screen, automation merely saves time. JS code is encapsulated in Python files under the scripts/ directory, invoked via browser-act --session <name> eval "$(python scripts/xxx.py {params})". $(...) is bash syntax; use the bash tool for execution. The eval token below refers to the browser-act CLI eval subcommand — always include the browser-act --session <name> prefix and the "$(...)" substitution.
Below are all atomic capabilities discovered and verified during the exploration phase, listed by command template with parameters. Simply invoke them as needed — no need to read scripts/*.py source code or re-verify. Only inspect scripts when execution fails for troubleshooting. Combine freely as needed during execution.
API: discover URLs from /llms.txt
Probes {origin}/llms.txt (a convention used by LLM-friendly documentation sites) and returns every URL it lists. Fast path — try this first.
On failure: {"error": true, "message": "No sitemap found at standard paths", "urls": []} — fall back to DOM link discovery.
DOM: discover URLs from the current page
Reads <a href> from the currently loaded page DOM, normalizes them to absolute URLs, drops fragments, asset extensions, and out-of-scope links, and optionally applies include / exclude glob filters. Use when llms.txt and sitemap are both unavailable, or to extend the queue with links discovered while crawling.
DOM: extract clean content + metadata + outbound links from the current page
The core extractor. Run this on every crawled page after wait stable. Returns the page body in the requested format(s), structured metadata, and the in-scope outbound links found on this page (so callers can extend the BFS queue without a second DOM pass).
positional start_url: scope URL — used to filter the outboundLinks array to in-scope links only.
--output-format: one of markdown, text, html, all. Default markdown. all includes every body field.
--remove-selectors: comma-separated CSS selectors to delete from the chosen content root before extraction (in addition to the built-in boilerplate list). Use this to strip site-specific chrome (e.g. .cookie-banner, #chat-widget).
--keep-selector: a single CSS selector identifying the main content area. If set, only this element's content is extracted (overrides the built-in content-root heuristic). Use this when the site has a known main wrapper, e.g. article.docs-content.
--include-globs / --exclude-globs: same semantics as discover-links; applied to the returned outboundLinks array.
Content-root heuristic (used when --keep-selector is not provided), in priority order: <main>, [role="main"], <article>, #content, .content, <body>.
{"error":false,"url":"https://example.com/docs/intro","crawl":{"loadedUrl":"https://example.com/docs/intro","loadedTime":"2026-06-25T04:37:23.643Z","referrerUrl":null},"metadata":{"canonicalUrl":"https://example.com/docs/intro","title":"Introduction — Example Docs","description":"Get started with Example.","author":null,"keywords":[],"languageCode":"en","publishedAt":null,"modifiedAt":null,"ogImage":"https://example.com/og.png","ogType":"website"},"text":"Introduction\n\nGet started with Example…","markdown":"# Introduction\n\nGet started with Example…","outboundLinks":["https://example.com/docs/install","https://example.com/docs/quick-start"]}
[AI Intervention] On pages with infinite scroll or lazy-loaded sections, before invoking this script: scroll down repeatedly (until page height stops growing or a max-scroll cap is hit) so the dynamic content is in the DOM. The script reads what is currently rendered — it cannot trigger lazy loading on its own.
Composite: full deep crawl from start URL(s)
End-to-end flow. The Agent orchestrates discovery → BFS queue → per-page extraction → persistence. Records are written one per page so that crashes can resume from where they stopped.
Step 1 — seed the queue:
For each start URL, perform discovery in this priority order and merge results. Stop discovery once the queue has enough URLs to honor max_pages.
a. eval "$(python scripts/discover-llms-txt.py '{origin}')" — instant full list when available.
b. eval "$(python scripts/discover-sitemap.py '{origin}' --max-urls {cap})" — broad coverage.
c. If both fail or return zero in-scope URLs: navigate to the start URL, wait stable, then eval "$(python scripts/discover-links.py '{start_url}' --include-globs '{globs}' --exclude-globs '{globs}')".
Filter all discovered URLs to scope: every URL must start with the start URL's origin + dirname/, and must satisfy include / exclude globs.
pages/ directory — one JSON file per successfully crawled page, named by URL hash
Step 3 — BFS loop (one URL at a time, in queue order, until max_pages reached or queue empty):
For each url popped from the queue:
a. Skip if url is in visited or its metadata.canonicalUrl (from a prior page) is already in visited.
b. navigate {url} → wait stable (use --timeout 60000 for slow sites).
c. (Optional, only when the target has lazy-loaded content) scroll down until height stable or 10 scrolls done.
d. eval "$(python scripts/extract-page-content.py '{start_url}' --output-format {format} --remove-selectors '{selectors}' --include-globs '{globs}' --exclude-globs '{globs}')".
e. If result error: true → record the failure into crawl_state.json#failed and continue. Do NOT retry blindly.
f. Write the JSON record to pages/{hash}.json.
g. Append url and metadata.canonicalUrl to visited.
h. For each link in outboundLinks: if not in visited and not already in queue, append to queue. Cap queue size at max_pages * 4 to bound memory.
i. Persist crawl_state.json after every page (resume on next run if interrupted).
Step 4 — finalize:
Emit a summary {total_pages, success_count, failed_count, duration_seconds, output_dir}. Optionally concatenate all pages/*.json into a single dataset.jsonl for downstream loading.
Configuration parameters (set by the Agent before Step 1 based on user request):
start_urls: list of seed URLs (one or more)
max_pages: hard cap on pages crawled, default 100
max_depth: hard cap on link depth from start URL, default unlimited (-1)
include_globs: JSON array, default []
exclude_globs: JSON array, default []
output_format: markdown | text | html | all, default markdown
remove_selectors: comma-separated site-specific selectors to strip, default ""
This is a recursive crawler, not a list with pages. Boundary control is by max_pages (queue length cap) and max_depth (links-away-from-start cap), not by API pagination. Termination: queue empty OR max_pages reached OR no more in-scope outbound links discovered.
Success Criteria
success_count >= 1 AND success_count / total_pages >= 0.8 AND extracted markdown body length per page > 100 chars for at least 80% of pages
Known Limitations
Pages behind authentication require the running browser to be logged in beforehand — this Skill does not handle login flows.
Pages whose content is rendered after async user interaction beyond simple scroll (e.g. clicking "Load more", expanding accordions to reveal content) need the Agent to add the relevant click before invoking extract-page-content; otherwise the hidden content will be missing.
<iframe> content is removed by default (treated as boilerplate). If a page's main content lives inside an iframe, the Agent must first navigate into the iframe URL and crawl it separately.
File downloads (PDF, DOCX, XLSX) are not handled. URLs with these extensions are intentionally filtered from the crawl queue.
Some sites' boilerplate is structurally indistinguishable from main content (e.g. "Was this page helpful?" footers placed inside <main>). The Agent should pass site-specific patterns via --remove-selectors to strip them.
Single-page applications that load content via JS after navigation may need a longer wait stable --timeout. Anti-scraping CAPTCHAs are not bypassed by this Skill; the calling browser must already pass them.
Execution Efficiency
Batch orchestration: Write a bash script to loop through the command templates serially within a single session; do not parallelize within one browser (prone to triggering anti-scraping restrictions). Refer to rate information in "Known Limitations" above to add appropriate intervals. To increase throughput, open multiple stealth browser sessions and distribute work across them — each session has an independent fingerprint so rate limits apply per session
Test before batch execution: After writing a batch script, you must first test with 1-2 items to verify the script runs correctly; only then run the full batch. Never skip testing and execute in batch directly
Reduce redundant pre-operations: When multiple steps depend on the same prerequisite state, complete them in batch under that state to avoid repeatedly establishing the same state
Error resumption: Save results item by item during batch processing; on failure, resume from the breakpoint rather than starting over
Prefer llms.txt / sitemap.xml over DOM discovery: A single fetch returns the full URL list; DOM discovery requires loading every page first. Always try the two API discovery routes first and only fall back to DOM when both return empty.
Polite delay between pages: Default to 500–1500 ms between page navigations to avoid burst patterns. Tighten only on sites you control or own.
Experience Notes
Path: {working-directory}/browser-act-skill-forge-memories/webcrawler-deep-crawl.memory.md (working directory is determined by the Agent running the Skill, typically the project root or current working directory)
Before execution: If the file exists, read it first — it records unexpected situations encountered during past executions (e.g., a strategy has become ineffective); adjust strategy order accordingly.
After execution: If an unexpected situation is encountered (strategy became ineffective, page redesigned, anti-scraping upgraded, better path discovered), append a line:
{YYYY-MM-DD}: {what happened} → {conclusion}
Normal execution does not write to the file. Do not record what keywords were used or how many results were returned — those are task outputs, not experience.
