Grok Search

Node scripts for web search, URL fetch and site mapping. Run npm install once in {baseDir}, the directory of this SKILL.md.

Choose The Script

• URL given → scripts/fetch.js.
• Site named, URL unknown → scripts/map.js, then fetch.js on the URLs you pick.
• Current/latest information, or URL unknown → scripts/search.js.
• Question about the user's local machine → inspect local files first; search only to interpret them.

Run the fewest commands that answer the question; run independent sub-questions in parallel. Do not chain map → fetch → search.

Commands

node {baseDir}/scripts/search.js "the question, with what to return"
node {baseDir}/scripts/search.js --instructions "format, language, what to leave out" "the question"
node {baseDir}/scripts/search.js --responses-allowed-domains github.com "the question"
node {baseDir}/scripts/search.js --source x --x-from-date 2026-07-01 "what people say about ..."
node {baseDir}/scripts/fetch.js https://example.com        # a cut page ends with the --offset to continue from
node {baseDir}/scripts/map.js https://docs.example.com --limit 20

Query rules:

• Only Grok reads the query, so write it as the question you want answered, with the fields you need: "What is the latest stable Go release, the full version number and its release date?" --instructions is optional, for how to answer (format, language, "quote and date only, no advice, say if not found"). Do not send a bag of keywords like GPT Codex 1M 272k context window; Grok fills the gaps with invented background.
• No code-search operators (repo:, path:, language:). Scope with --responses-allowed-domains.
• One search per sub-question is usually enough; to check a claim, fetch its source instead of searching again. Search again only for a gap you can name. Do not carry the previous round's conclusion words into the new question; for the same X handle, change the gap, not the phrasing.

Search Source

Default is web. --source x for what people are saying, a named account's posts, a thread, or a claim only circulating on X. --source both for a current event where reporting and reaction both matter. --source x cannot fall back to the web.

• For "is it available now / current state" questions, start with --x-from-date 60–90 days back; widen only if empty (diagnostics.warnings says when the window read no posts). Old issues explain history, not the present.
• X posts are personal statements: cite handle and date (both live in answer.text, not the card), keep claims separate from confirmations, confirm facts at the official source.
• Cost: when an X search reports responses_tool_calls.total above 6, add --responses-parallel-tool-calls false next time: one call per turn, so max_turns caps the count; below 6 it saves nothing.
• Handle filters, dates, media flags, card fields: references/x-search.md.

Fetch Cost

• X posts: auto tries Direct first (free, main post with date). --provider firecrawl only for the thread or replies; about 30 credits per post.
• Fetch only evidence that would change the conclusion: one or two URLs per turn, a representative post or two for experience reports, never re-fetch text you have. Reddit and YouTube return shells from every provider. WeChat articles (mp.weixin.qq.com) need --provider deepseek or --provider firecrawl; the others get a verification page.
• The same URL again within 15 minutes comes from a local cache (no provider call); a different --provider or --no-cache fetches anew.
• When citing decisive evidence, record date, model version, client, and login method.

Reading Results

Every script prints one JSON object, also on failure (non-zero exit, stderr line).

• error: read error.message, error.code, diagnostics.provider_attempts. Change something before retrying (query, --provider, --model). DEADLINE_EXCEEDED: the whole command hit its time budget (default 240s, --deadline N).
• A line starting with [grok-search] inside content.text or answer.text changes how to read it: the text was cut (the line says where the rest is), the page redirected to the site's home page, or the provider got no text. diagnostics.warnings repeats these and adds the rest: providers skipped (e.g. Firecrawl in cooldown) or failed, a relay serving another model than requested. diagnostics.provider_attempts lists each provider tried.
• Search: answer.text, then sources.items (up to 12 of Grok's sources, cited ones kept first, so a cited link may appear only in answer.text). source_type citation = used in the answer, searched = only listed; opened: true = Grok read the page. sources.raw_path always points to the run record (full list, answer, tool calls); read it in chunks only when the cards are not enough. Check diagnostics.cost_usd and diagnostics.search_budget (advisory budget vs. calls made).
• Fetch: content.text, up to --max-chars characters (default 12000) from --offset (default 0); content.chars counts page text only. A cut page ends with a [grok-search] line giving the next --offset and content.full_path: continue from there instead of raising --max-chars. stdout stays under 45KB whatever --max-chars says. Non-HTML files come back as served (JSON re-indented). metadata has title/author/published_at when the provider had them. diagnostics.run_path is the run record.
• Map: urls; fetch the few you need. A URL with a path maps that path and below, on that site only; map the site root for the whole site.

Provider order, proxy, cooldown, config errors: references/providers.md. Multi-part or conflicting research: references/planning.md first.