Web Research

Produce a thorough, cited answer to a research question by searching the web, reading the best sources, and synthesizing a report — with the heavy reading done in a background task so the main conversation stays clean.

Question to research: $ARGUMENTS

How to run it

1. If the question in $ARGUMENTS is empty or vague, ask the user for a clear research question (and scope/time-frame) before proceeding.
2. Pick a short kebab-case report name from the topic, e.g. research-<topic>.
3. Delegate the work to a background task with internal__run_background so the many fetched pages never enter this conversation's context. Pass:
   • description: "Web research: <topic>"
   • save_to_report: the report name from step 2
   • max_time_seconds: 600 (research is multi-step)
   • prompt: the research brief below, with $ARGUMENTS filled in.
4. Tell the user the research is running in the background, name the report it will write, and that they'll be notified when it completes. Do not start fetching pages yourself in this conversation.
5. When the task completes, read the report with internal__read_report and give the user a short summary plus the report name. If they want more, point them at the full report.

If background tasks are unavailable (e.g. non-interactive mode), run the research brief below inline instead, then save the result with internal__write_report.

Research brief (the background prompt)

You are researching: $ARGUMENTS

1. Use internal__web_search to find sources. Start broad, then run 2–4 refined queries covering sub-questions and alternative phrasings. Prefer primary and authoritative sources.
2. Use internal__web_fetch to read the full content of the ~5–8 most relevant results (not just snippets). For pages that fail to fetch, try another source.
3. Cross-check claims across at least two independent sources. Note where sources disagree, and flag anything you could not verify.
4. Synthesize a structured markdown report:
   • Summary — the direct answer in a few sentences.
   • Findings — organized by sub-topic, with inline citations like [1].
   • Caveats / open questions — gaps, disagreements, recency limits.
   • Sources — numbered list of [n] Title — URL for every citation used.
5. Ground every non-obvious claim in a citation. Do not invent URLs or facts; if the web results don't answer part of the question, say so explicitly.
6. Save the finished report (this happens automatically via save_to_report).

Notes

• Works with no API keys (DuckDuckGo search + built-in HTTP fetch). Relevance and anti-bot coverage improve automatically when Tavily / Brave / Firecrawl keys are set — this skill is provider-agnostic.
• Use __save_to_file on internal__web_fetch for very large pages so they don't bloat even the background task's context.