chrome-devtools-site-search

Chrome DevTools Site Search

Contract

Prereqs:

• mcp_servers.chrome-devtools must be available and connectable.
• User-provided site scope (domain / URL prefix) and a concrete query.

Inputs:

• Site scope (domain / allow-list / URL prefix).
• Query terms and constraints; optional open policy (open 1 vs top N).

Outputs:

• A result summary backed by observed page content (real titles/URLs).
• Matching pages opened in the MCP-controlled browser session for user review.

Exit codes:

• N/A (interactive MCP workflow; no repo scripts)

Failure modes:

• MCP server missing/unavailable/timeouts (must stop and report; no fallback scraping).
• Site blocks automation (consent wall/login/bot protection) or content not extractable reliably.

Purpose (Parsing + Automation Feasibility)

• This skill’s workflow is primarily for testing whether a site can be reliably parsed (content extraction) and whether automated browsing flows are feasible using a real browser controlled by mcp_servers.chrome-devtools.

Setup

• Use the MCP server mcp_servers.chrome-devtools to control a real browser session (tabs, navigation, interaction, and page content extraction).
• If mcp_servers.chrome-devtools is not available (missing, cannot start, cannot connect, times out), STOP immediately and report the failure and what needs to be enabled/configured. Do not fall back to non-browser scraping or alternative search methods outside the MCP-controlled browser.

Inputs

Collect (or infer) these inputs from the user request:

• Site scope: a domain, URL prefix, or an explicit allow-list of domains/paths.
• Query conditions: keywords and any constraints (title must contain X, language, timeframe, section/category, content type).
• Open policy: whether to open a single result (“any one”) or multiple results (“top N”).

Ask at most 1-2 clarifying questions if the request is underspecified. Defaults:

• If the user does not specify a count, open up to 3 results.
• Cap auto-open to 5 pages unless the user requests more.

Workflow

1. Confirm scope and intent

   • Restate the site scope and query conditions in one sentence.
   • If the user’s terms are ambiguous, ask for the exact keyword(s) and the desired count of pages to open.

2. Navigate to the site

   • Open the site entry page (home, section page, or known search page).
   • Keep a dedicated “results” tab whenever possible.

3. Choose a search strategy (prefer in-site search)

   • Prefer the site’s native search UI or search endpoint.
   • If no native search is discoverable, use a reputable external search engine with a strict site restriction (for example, site:<domain> <keywords>), and only open results that match the requested scope.

4. Collect candidate results with evidence

   • For each candidate, capture:
     • Title text as displayed on the site
     • Final URL after navigation/redirects (canonical when available)
     • Date/time if visible (otherwise record “date not visible”)
     • A 1-sentence gist from the page’s lede/summary
     • Why it matches the user’s condition(s)
     • (Feasibility) Whether the title/date/lede were extractable via stable DOM selectors, and any automation blockers (e.g., consent wall, login, bot protection, infinite scroll)
   • Do not rely on search snippets alone; open the page and verify the title/lede against the conditions.

5. Summarize results

   • Summarize in a compact list (3-6 bullets max unless requested), each containing: Title — Date (if known) — 1-sentence gist — URL.
   • If there are no matches, state that clearly and suggest the smallest query refinement (different keyword, broader timeframe, alternate spelling).

6. Open result pages for the user

   • Open the number of pages implied by the user’s open policy (or defaults).
   • Prefer opening pages in background tabs so the results list remains available.
   • Leave all opened pages/tabs in place for the user; do not close or “clean up” tabs.

Quality and safety rules

• Do not fabricate results. Every summarized item must be backed by an observed page and a real URL obtained during the run.
• Stay within the user-provided site scope. If a result is off-scope but highly relevant, ask before opening it.
• If the user request involves disallowed content, refuse that part and offer a safe alternative (for example, summarize high-level workflow without visiting the site).