web-scraping

name	web-scraping
description	Extract and summarize content from the currently opened web view with robust fallback.
metadata	{"annolid":{"always":true}}

Use this skill when the user asks about the currently opened web page, says "this page/current page", or asks for web extraction/summarization.

Primary goal: return grounded page content quickly, with source links and minimal hallucination.

Workflow

1. Check active page state first.

• Use gui_web_get_state().
• If no active page, run a web lookup flow (gui_web_run_steps or web_search) and then continue.

2. Extract page text from the active web view.

• Use gui_web_get_dom_text(max_chars=9000) first.
• Prefer gui_web_extract_structured(fields=[...]) for schema-shaped answers.
• When field-specific regex is brittle, provide selector_hints (CSS-like or keyword hints) and set extraction_mode="hint" for resilient line-level extraction.
• If content is too short/noisy, use gui_web_run_steps with get_text after a short wait and optional scroll.
• If visual context matters, use gui_web_capture_screenshot(max_width=1600) and then gui_web_describe_view(max_width=1600).

3. Fallback to scraper/search tools when needed.

• Use web_search(query, count) (Scrapling-first backend).
• Use web_fetch(url, extractMode="text") when a target URL is known.
• Prefer returning the best available grounded result rather than refusal text.

4. Response style.

• Return concise extracted facts.
• Include source URL(s).
• If data is time-sensitive (weather, prices, news), clearly label it as current lookup output.

Reliability rules

• If model output is empty or uncertain, execute tool fallback immediately.
• Do not claim browsing is unavailable before attempting available web tools.
• Prioritize the currently open page when the user references "this/current/open page".