researching-web - SKILL.md Agent Skill

name: researching-web description: | Researches questions using web search and generates reports with confidence scores and sources. Use when: (1) User asks a question requiring current information, (2) User says "find out about X", (3) User gives URL to extract data from, (4) Question about code/API/library, (5) Comparing options. Triggers: "/research", "find out about", "research this".

Web Research

Orchestrates multiple search tools for hybrid search, parallel extraction, and structured output.

Pipeline

Query -> Classify -> Search (parallel) -> Score -> Extract (parallel) -> Synthesize -> Verify -> Output

Tools: Uses available web search and extraction tools (Exa, Tabstack, WebSearch, WebFetch, or similar MCP tools).

Progress Format

Show progress with real data at each step:

======================================
   RESEARCHING: "{query}"
======================================
[##----] Search: Source A 8 | Source B 5 -> 10 unique

[###---] Scoring: domain1.com (82), domain2.com (78), ...
         Selected: 5 sources above threshold

[####--] Extracting (parallel)...
         + domain1.com
         + domain2.com
         ...

[#####-] Synthesizing + Verifying...

[######] Confidence: 85%
         |-- Consensus: 4/5 agree
         |-- Top source: 82 (official docs tier)
         |-- Freshness: 5/5 from current year
         |-- Contradictions: none

RESEARCH DEPTH
   Pages analyzed: 12 | Facts extracted: 47
   Sources: 5 (2 official, 2 research, 1 blog)
   Coverage: High, multiple independent confirmations
======================================

Step 1: Classify and Plan

Type	Signals	Sources	Search Strategy
Fact	"what is", "when", "how much"	1-2	Single search tool
How-to	"how to", tutorial	2-3	Single search tool
Comparison	"vs", "compare", "best"	5+	Hybrid (multiple tools)
Overview	"explain", "tell me about"	3-5	Hybrid
Code/API	library, SDK, docs	1-2	Code-specific search

Step 2: Search

Has URL: skip to extraction.

No URL: search with query augmentation (synonyms, English version for tech).

For hybrid: call multiple search tools in a single message for parallel execution.

Fallback: If a tool is unavailable, note it and continue with remaining tools.

Step 3: Score and Select

Trust Claude's judgment. Prefer: official docs > research > GitHub > Stack Overflow > blogs > forums.

Skip: SEO spam, paywalls, content older than 2 years (for tech).

Step 4: Extract (Parallel)

Call ALL extractions in a single message for parallel processing.

If extraction fails: note the failure and continue with others.

Step 5: Synthesize

Merge facts, deduplicate
Note contradictions with source attribution
Identify consensus vs disputed points

Step 6: Verify and Detect Contradictions

Before finalizing, actively look for contradictions between sources. When found, display:

CONTRADICTION DETECTED:
   Source A (domain.com): "Claim X"
   Source B (other.com): "Claim Y"
   -> Likely cause: {different timeframes / methodology / scope}

Also check:

Single-source claims: mark as "unverified"
Non-authoritative sources for topic: adjust confidence

Step 7: Output (Zero Friction)

Auto-select format, no questions asked:

"vs"/"compare": Comparison Table HTML
Simple question: Answer in chat + sources
Complex topic: Full HTML Report

Generate immediately. If the user wants a different format, they will ask.

Include in every output:

Confidence score with breakdown
Contradictions found (if any)
Research depth stats

Fallback

Tool fails: inform, continue with remaining. Search empty: simplify query. All weak: suggest refining.