Researching Web

Turn any question into a research report with confidence scores, contradiction detection, and source attribution.

Features • Demo • Install • How It Works • Examples

---

Why This Exists

Most AI assistants give you answers. This skill gives you researched insights with:

Traditional AI	Researching Web
Single response	Multi-source synthesis
"Trust me"	Confidence score with breakdown
Hidden sources	Full attribution + scoring
Ignores conflicts	Detects contradictions
Text dump	Beautiful HTML reports

---

Features

• Hybrid Search — Combines Exa semantic search + Tabstack for comprehensive coverage
• Parallel Extraction — Fetches 5+ sources simultaneously for speed
• Contradiction Detection — Finds and explains conflicting information between sources
• Confidence Scoring — Transparent breakdown: consensus, authority, freshness
• Research Depth Stats — Shows pages analyzed, facts extracted, coverage level
• Zero Friction — Auto-selects best output format, no questions asked
• Beautiful HTML Reports — Dark/light theme, responsive, screenshot-worthy

---

Demo

Terminal output:

Generated HTML report:

See full terminal output

══════════════════════════════════════════
   RESEARCHING: "AI coding tools market 2026"
══════════════════════════════════════════
[■■□□□□] Search: Exa 8 | Tabstack 5 → 10 unique

[■■■□□□] Scoring: cbinsights.com (90), venturebeat.com (85)...
         Selected: 5 sources above threshold

[■■■■□□] Extracting (parallel)...
         ✓ cbinsights.com
         ✓ venturebeat.com
         ✓ digitalapplied.com
         ✓ index.dev
         ✓ secondtalent.com

[■■■■■□] Synthesizing + Verifying...

⚠️ CONTRADICTION DETECTED:
   CB Insights: "Market size $4B"
   SecondTalent: "Market $4.91B in 2024"
   → Likely cause: different scope definitions

[■■■■■■] Confidence: 87%
         ├─ Consensus: 4/5 agree
         ├─ Top source: 90 (CB Insights)
         ├─ Freshness: 5/5 from 2024-2026
         └─ Contradictions: 2 found

📊 RESEARCH DEPTH
   Pages analyzed: 5 | Facts extracted: 64
   Sources: 5 (1 premium, 1 news, 3 industry)
   Coverage: High
══════════════════════════════════════════
✓ READY: insight-ai-coding-market-2026.html

---

Installation

This is a Claude Code skill. To install:

# Clone to your skills directory
git clone https://github.com/vasilievyakov/researching-web-skill.git \
  ~/.claude/skills/researching-web

Requirements

You need these MCP servers configured in Claude Code:

MCP Server	Purpose	Required
Exa	Semantic web search	Yes
Tabstack	Content extraction	Recommended

MCP Configuration Example

{
  "mcpServers": {
    "exa": {
      "command": "npx",
      "args": ["-y", "@anthropic/mcp-exa"],
      "env": { "EXA_API_KEY": "your-key" }
    },
    "tabstack": {
      "command": "npx",
      "args": ["-y", "@anthropic/mcp-tabstack"],
      "env": { "TABSTACK_API_KEY": "your-key" }
    }
  }
}

---

How It Works

┌─────────────────────────────────────────────────────────────────┐
│  Query                                                          │
│  "Compare React vs Vue for enterprise 2026"                     │
└─────────────────┬───────────────────────────────────────────────┘
                  │
                  ▼
┌─────────────────────────────────────────────────────────────────┐
│  1. CLASSIFY                                                    │
│  Type: Comparison → 5+ sources, hybrid search                   │
└─────────────────┬───────────────────────────────────────────────┘
                  │
                  ▼
┌─────────────────────────────────────────────────────────────────┐
│  2. SEARCH (parallel)                                           │
│  ┌──────────────┐    ┌──────────────┐                          │
│  │ Exa Semantic │    │  Tabstack    │  → Dedupe → 10 unique    │
│  │   8 results  │    │  5 results   │                          │
│  └──────────────┘    └──────────────┘                          │
└─────────────────┬───────────────────────────────────────────────┘
                  │
                  ▼
┌─────────────────────────────────────────────────────────────────┐
│  3. SCORE & SELECT                                              │
│  official docs (90) > research (80) > blogs (60) > forums (40) │
│  → Top 5 sources selected                                       │
└─────────────────┬───────────────────────────────────────────────┘
                  │
                  ▼
┌─────────────────────────────────────────────────────────────────┐
│  4. EXTRACT (parallel)                                          │
│  All 5 sources extracted simultaneously via Tabstack            │
└─────────────────┬───────────────────────────────────────────────┘
                  │
                  ▼
┌─────────────────────────────────────────────────────────────────┐
│  5. SYNTHESIZE                                                  │
│  Merge facts, dedupe, identify consensus vs disputed            │
└─────────────────┬───────────────────────────────────────────────┘
                  │
                  ▼
┌─────────────────────────────────────────────────────────────────┐
│  6. VERIFY                                                      │
│  ⚠️ Detect contradictions between sources                       │
│  ⚠️ Flag single-source claims as "unverified"                   │
│  → Calculate confidence score                                   │
└─────────────────┬───────────────────────────────────────────────┘
                  │
                  ▼
┌─────────────────────────────────────────────────────────────────┐
│  7. OUTPUT                                                      │
│  Auto-select format → Generate HTML report                      │
│  insight-react-vs-vue-2026.html                                 │
└─────────────────────────────────────────────────────────────────┘

---

Examples

Query Types & Output

Query Type	Example	Auto-Output
Comparison	"Claude Code vs Cursor vs Windsurf"	Comparison table HTML
Fact	"What is Claude's context window?"	Quick answer in chat
Overview	"AI coding tools market 2026"	Full HTML report
How-to	"How to set up MCP servers"	Step-by-step guide

Sample Reports

• AI Coding Tools Market 2026 — Market analysis with contradiction detection
• Claude Code vs Cursor vs Windsurf — Tool comparison

---

Comparison vs Alternatives

Feature	This Skill	ChatGPT Search	Perplexity
Multi-source synthesis	✅ 5+ sources	✅	✅
Source scoring	✅ Transparent	❌ Hidden	⚠️ Partial
Contradiction detection	✅ Unique	❌	❌
Confidence breakdown	✅ Detailed	❌	⚠️ Simple
HTML report output	✅ Beautiful	❌	❌
Research depth stats	✅ Pages/facts	❌	❌
Customizable	✅ Edit skill	❌	❌
Works offline*	✅ With local MCP	❌	❌

*With self-hosted search MCP servers

---

Configuration

The skill auto-detects query type and selects the best approach:

Query Signal	Sources	Strategy
"vs", "compare"	5+	Hybrid search, comparison table
"what is", "when"	1-2	Exa only, quick answer
"how to"	2-3	Exa only, step-by-step
Complex topic	3-5	Hybrid, full report
Code/API question	1-2	get_code_context_exa

---

Output Formats

HTML Report (default for complex queries)

• Responsive dark/light theme
• Confidence gauge with breakdown
• Contradiction boxes with explanations
• Research depth visualization
• Clickable source links with scores

Quick Summary (for simple questions)

• Answer directly in chat
• Sources listed with scores
• No file generated

---

Learnings

Why Contradiction Detection?

Most AI search tools hide conflicting information — they pick one answer and present it as truth. But real research often reveals disagreements between sources.

Example: When researching "AI coding tools market size", we found:

• CB Insights: "$4B market"
• SecondTalent: "$4.91B market"

Without contradiction detection, you'd get one number and trust it. With it, you see both — and understand why they differ (scope definitions). This is the difference between an answer and an insight.

From 556 to 127 Lines

The first version of this skill was 556 lines — detailed explanations, scoring tables, step-by-step instructions.

The problem: Claude is already smart. Over-explaining makes the skill slower and harder to maintain.

The fix: Applied "Andrej Karpathy's rule" — trust the model, remove everything it already knows. Keep only:

• What tools to use (MCP names)
• What output format to show
• What makes this skill unique (contradictions, confidence)

Result: 127 lines that work better than 556.

---

Contributing

Contributions welcome! See CONTRIBUTING.md.

Ideas for improvement:

• Add more output formats (Markdown, PDF)
• Support more MCP search providers
• Multilingual query support
• Custom confidence thresholds

---

License

MIT License — see LICENSE for details.

---

_{Built with Claude Code • Powered by Exa + Tabstack MCP}