Web Research

This skill guides systematic investigation and synthesis of information from external web sources. The user poses a research question and optionally specifies scope constraints (date range, domain preferences, citation format). Apply the workflow below, then return synthesized findings alongside a credibility audit log.

🎯 When to Use

• User asks to find, verify, compare, or summarize information from external sources
• User requests citations, fact-checking, competitive analysis, or literature reviews
• Query involves time-sensitive data, emerging topics, or claims requiring external validation
• Do NOT use when the user wants to generate original code, creative content, or hypothetical scenarios — refer them to appropriate generation workflows
• Do NOT use for questions answerable from internal knowledge alone without external verification needs — avoid unnecessary search overhead
• Do NOT use for queries requiring authenticated/paywalled content access, private databases, or real-time APIs unless the user explicitly provides credentials and authorizes external tool use

🧠 Core Workflow

Step 1 — Clarify & Scope Parse the research question to identify:

• Core claim or information need (factual verification, trend analysis, comparison)
• Temporal constraints (e.g., "last 12 months", "post-2020 studies")
• Domain preferences or exclusions (e.g., "peer-reviewed only", "exclude blogs")
• Output format requirements (APA, MLA, inline URLs, bullet summary) If scope is ambiguous (e.g., "tell me about AI"), ask one targeted clarifying question before proceeding. Never silently assume breadth or depth preferences.

Step 2 — Query Strategy & Source Discovery Generate 3-5 targeted search queries using advanced operators:

• site: for domain restriction (e.g., site:.gov, site:arxiv.org)
• filetype: for document type filtering (e.g., filetype:pdf for reports)
• intitle:/inurl: for precision targeting
• Date-range filters where supported Execute searches and collect top 10-15 candidate sources per query. Deduplicate by URL canonicalization and content hashing.

Step 3 — Credibility Assessment & Extraction Rules Evaluate each source using a weighted rubric. Log scores and rationale.

1. Authority — Prefer .gov, .edu, peer-reviewed journals, established news outlets with editorial standards. Flag personal blogs, anonymous sources, or domains with known misinformation histories with ⚠️ low-authority.
2. Recency — Prioritize sources within user-specified timeframe. For time-sensitive topics (tech, medicine, policy), flag sources >2 years old unless seminal/historical context is requested.
3. Bias & Transparency — Assess political/commercial bias using domain reputation databases. Prefer sources with clear methodology, funding disclosures, and correction policies. Flag opinion pieces as opinion not fact.
4. Corroboration — Cross-check key claims across ≥2 independent high-credibility sources. Flag singleton claims (only one source) with ⚠️ uncorroborated.
5. Extraction — Pull direct quotes, statistics, or conclusions with precise attribution (author, publication date, URL, access date). Never paraphrase in a way that alters original meaning or certainty level.

Step 4 — Synthesis & Gap Analysis Group findings by theme or claim. Resolve conflicts by weighting source credibility.

• If sources agree: synthesize consensus with representative citations
• If sources conflict: present both positions with credibility context, avoid false equivalence
• Identify gaps: missing perspectives, outdated data, or unanswered sub-questions Document all synthesis decisions to enable auditability.

Step 5 — Return Output Provide structured JSON report followed by optional human-readable summary. Always include credibility indicators, citation metadata, and warnings for low-confidence findings. Never present unverified claims as established fact.

📋 Output Format

Return two sections: Section 1 — Research Report

{
  "status": "success | partial | error",
  "metadata": {
    "research_question": "string",
    "queries_executed": ["query1", "query2"],
    "sources_evaluated": 0,
    "sources_cited": 0,
    "time_range_covered": "YYYY-MM-DD to YYYY-MM-DD",
    "domains_covered": [".gov", ".org"]
  },
  "findings": [
    {
      "claim": "Concise statement of finding",
      "confidence": "high | medium | low",
      "corroboration_count": 2,
      "sources": [
        {
          "title": "Article Title",
          "author": "Author Name or Organization",
          "url": "https://...",
          "published_date": "YYYY-MM-DD",
          "accessed_date": "YYYY-MM-DD",
          "credibility_score": 0.92,
          "credibility_factors": ["peer_reviewed", "recent", "transparent_methodology"],
          "bias_flag": "none | left | right | commercial | opinion",
          "excerpt": "Direct quote or precise paraphrase with page/section"
        }
      ]
    }
  ],
  "summary": "Executive synthesis: X high-confidence findings, Y contested claims. Key consensus: ...",
  "warnings": [
    {
      "claim": "string",
      "issue": "Description of credibility concern, data gap, or conflicting evidence",
      "sources_involved": ["url1", "url2"],
      "suggestion": "Manual verification recommended; consider consulting [alternative source type]"
    }
  ],
  "gaps": [
    "Unanswered sub-question or missing perspective requiring further research"
  ]
}

Section 2 — Human-Readable Summary (Optional) If requested, provide a markdown-formatted summary with inline citations. Use [Author, Year] or [URL] format per user preference. Add > ⚠️ callouts for low-confidence findings or contested claims.

⚠️ Fallback Behavior

If the research question is too broad, ambiguous, or lacks actionable scope, respond with:

{
  "status": "error",
  "message": "Research query requires clarification to execute effectively.",
  "suggestion": "Please specify: (1) the exact claim or topic to investigate, (2) preferred date range, and (3) any domain preferences (e.g., academic, government, news)."
}

If no credible sources are found after exhaustive search, return:

{
  "status": "partial",
  "message": "No high-credibility sources found matching query constraints.",
  "attempted_queries": ["..."],
  "suggestion": "Consider broadening date range, relaxing domain filters, or rephrasing the research question."
}

Never silently present low-credibility sources as authoritative. If the user requests real-time data beyond the model's knowledge cutoff, explicitly flag this limitation and request authorization to use live search tools before proceeding.