web-research

# Web Research ## Purpose Conduct rigorous, multi-step web research that goes beyond a single search query. This skill orchestrates a structured research workflow: planning a search strategy, executing multiple targeted queries, cross-verifying claims across independent sources, and synthesizing findings into a coherent, well-sourced report. ## Guiding Principles 1. **Breadth before depth** — Cast a wide net with diverse queries before drilling into any single source. Reformulate queries when initial results are thin. 2. **Source triangulation** — Never trust a single source. Every key claim must appear in at least two independent sources before it is treated as established. 3. **Recency awareness** — Prefer recent sources for fast-moving topics. Flag when the most recent source is older than the topic warrants. 4. **Bias detection** — Note when sources share a common owner, funding body, or obvious editorial slant. Weight accordingly. 5. **Transparency** — Every factual statement in the final report must link back to the source URL it was derived from. ## Workflow ### Phase 1 — Understand the Question Before any search, decompose the user's request: - Identify the **core question** (what must be answered). - List **sub-questions** that feed into the core answer. - Note any **constraints**: date range, geography, domain, language. - Decide on a **search budget**: how many queries are proportionate to the task (typically 3–8 queries; simple factual lookups need fewer). Output a brief research plan (≤ 6 bullet points) and confirm with yourself before proceeding. ### Phase 2 — Execute Searches For each sub-question, craft a targeted search query: 1. Use **specific, keyword-rich** queries — avoid full sentences. 2. Vary phrasing across queries to surface different result sets. 3. Include **domain-specific qualifiers** when useful (e.g., `site:arxiv.org`, `filetype:pdf`, `"exact phrase"`). 4. After each `web_search` call, scan the snippets. If a result looks authoritative or contains the needed data, fetch the full page with `web_fetch`. 5. Record each source: URL, title, date (if available), and a 1–2 sentence summary of what it contributes. Do NOT fetch every result — only those that plausibly advance an answer. ### Phase 3 — Cross-Verify Lay out the key claims you have gathered and check: - Does each claim appear in **≥ 2 independent sources**? - Are any claims **contradicted** by another source? If so, note the contradiction and assess which source is more credible (recency, authority, methodology). - Are there **gaps** — sub-questions that no source adequately addresses? If so, run 1–2 additional targeted searches. ### Phase 4 — Synthesize Produce a structured report: ``` ## Summary <2–4 sentence executive summary> ## Key Findings 1. <Finding with inline source reference [1]> 2. ... ## Contradictions / Uncertainties - <Any unresolved conflicts or low-confidence claims> ## Sources [1] <title> — <url> (accessed <date>) [2] ... ``` Rules for the report: - Lead with the answer — do not bury it after methodology. - Use numbered source references; every factual claim must cite at least one. - Clearly separate established facts from speculation or single-source claims. - If the research is inconclusive, say so explicitly rather than hedging. ### Phase 5 — Self-Check Before delivering: - Re-read the original question. Does the report actually answer it? - Are all source URLs valid (no hallucinated links)? - Is the report length proportionate to the question's complexity? ## Error Handling | Situation | Action | |---|---| | `web_search` returns no results | Reformulate the query with broader or alternative terms; try up to 2 reformulations before reporting the gap. | | `web_fetch` fails or returns empty | Note the URL as inaccessible; do not cite it. Rely on the search snippet if it contained the needed fact. | | Contradictory sources with equal credibility | Present both perspectives explicitly; do not silently pick one. | | Topic is too broad | Ask the user to narrow scope before consuming the full search budget. | ## Limitations - This skill relies on `web_search` and `web_fetch` tools. If they are unavailable, the skill cannot execute. - Results are bounded by what the search engine indexes; paywalled or dynamically rendered content may be inaccessible. - Always note when a topic requires expertise that web search alone cannot reliably provide (medical, legal, financial advice).