web-research

Systematic multi-source web research with verification, citation, and contradiction detection. Use when gathering facts from the web, verifying claims across sources, or building research briefs.

Model	Source
sonnet	pack: recon

Full Reference

web-research

┏━ 🔍 web-research ━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━┓
┃ Systematic multi-source research with           ┃
┃ verification, citation, contradiction detection ┃
┗━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━┛

Multi-source research pipeline — decompose queries, search, fetch, cross-validate, flag confidence, and deliver a citable brief. No claim ships without at least 3 sources. No statistic ships without 5.

Item	Value
Phases	5 (Decompose → Search → Fetch → Verify → Output)
Min sources per claim	3
Min sources per statistic	5
Confidence levels	high / medium / speculative
Output	RESEARCH_BRIEF.md with inline citations
Contradiction behavior	Surface all, don’t resolve silently

Reference Index

I want to…	File
Decompose queries, use date-aware patterns, filter by domain	reference/search-strategies.md
Score source credibility, cross-validate, detect contradictions	reference/source-verification.md
Coordinate WebFetch + WebSearch, extract structured data	reference/extraction-patterns.md
Format RESEARCH_BRIEF.md, citations, confidence tags	reference/output-format.md

When to Activate

Trigger	Example
Fact-finding request	”What are the current Stripe fee tiers?”
Claim verification	”Is it true that X framework is deprecated?”
Multi-source research	”Research competitor pricing for SaaS email tools”
Statistics gathering	”What’s the average checkout abandonment rate?”
Background research	”Get context on this company before the call”

Do NOT activate for: single-file reads, code questions with no web dependency, or tasks where all needed data is already in context.

━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━

Phase 1 — Decompose

Break the research goal into atomic questions before searching.

Research goal: "What is the current market share of Tailwind CSS vs Bootstrap?"

Atomic questions:
1. What is Tailwind CSS current adoption/usage percentage?
2. What is Bootstrap current adoption/usage percentage?
3. What surveys or studies cover CSS framework market share?
4. What is the most recent data (year) available?
5. Are there contradictions between survey sources?

Rules:

• Never search the raw goal — always decompose first
• Each atomic question maps to one or more search queries
• Flag which questions require statistics (min 5 sources each)
• Flag which questions are time-sensitive (add date filter)

See reference/search-strategies.md for query construction patterns.

Phase 2 — Search

Run WebSearch for each atomic question using decomposed queries.

Query pattern: "<topic> <qualifier> <year>"
Example: "Tailwind CSS adoption rate survey 2026"

Rules:

• Minimum 2 query variants per atomic question
• Use date-aware queries for anything time-sensitive
• Filter by domain type when authority matters (.gov, .edu, primary sources)
• Log every search query and its result count
• Stop at 15 searches total — if still unsatisfied, escalate to user

See reference/search-strategies.md → Iterative Deepening.

Phase 3 — Fetch

Use WebFetch to extract detail from top-ranked results. When WebFetch fails, fall back to firecrawl.

Rules:

• Fetch only the top 3 URLs per atomic question
• Extract key facts, quotes, dates, and data points
• Note publication date and author/publisher for every source
• Skip paywalled, login-required, or dead URLs — log as ✗ inaccessible
• Never summarize without reading — always fetch before citing

Firecrawl Fallback

When WebFetch fails (paywall, auth wall, timeout, 403, 429), try firecrawl before marking a source as inaccessible:

WebFetch Result	Action
200 OK	Use WebFetch content — no fallback needed
403 / paywall / auth	Try firecrawl scrape on the URL
429 rate limited	Back off 4s, retry WebFetch once, then firecrawl
Timeout (>30s)	Try firecrawl scrape on the URL
404 / gone	Skip — no fallback

Firecrawl budget: max 5 firecrawl pages per research session (in addition to the 15-fetch WebFetch budget).

See reference/extraction-patterns.md for WebFetch + firecrawl coordination.

Phase 4 — Verify

Cross-validate every claim before including it in the brief.

Claim type	Min sources	Confidence threshold
Factual statement	3	2/3 agree → high
Statistic / number	5	3/5 agree within 10% → high
Opinion / trend	2	Cite all, label speculative
Breaking news	3	Check publication date lag

Contradiction protocol:

1. Surface the contradiction explicitly — never pick a winner silently
2. Note which sources disagree and why (methodology, date, scope)
3. Label the claim [CONTRADICTED] in the brief
4. Recommend the most authoritative source, but show all

Credibility scoring and source-type rankings are in reference/source-verification.md.

Phase 5 — Output

Produce RESEARCH_BRIEF.md following the standard template.

Every fact gets:

• Inline citation: [1]
• Confidence tag: [HIGH] / [MEDIUM] / [SPECULATIVE]
• Source entry in bibliography with URL + date accessed

Template, citation format, and confidence tagging rules are in reference/output-format.md.

━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━

Confidence Levels

Level	Meaning	Requirement
[HIGH]	Verified across multiple credible sources	3+ independent sources agree, no contradictions
[MEDIUM]	Supported but not fully corroborated	2 sources agree, or 1 authoritative primary source
[SPECULATIVE]	Single source, unclear methodology, or unverifiable	Cite source, flag explicitly

Always show confidence inline. Never omit it.

Common Mistakes

Mistake	Correct approach
Citing a search result without fetching the page	Always WebFetch before citing
Averaging contradictory statistics	Surface the contradiction, cite all values
Treating .com blog posts as primary sources	Score credibility first — see source-verification.md
Using data older than 2 years without flagging	Mark as [OUTDATED — YYYY] in the brief
Stopping at 1 source and calling it done	Min 3 per claim — no exceptions
Resolving contradictions silently	Always surface, always label [CONTRADICTED]
Including unsearchable claims as speculative without note	Flag [UNVERIFIABLE] and state why

Tool Coordination

WebSearch  →  URL discovery, result scanning, quick fact checks
WebFetch   →  Deep extraction from specific pages, quote pulling

Never use WebFetch cold on a URL you haven’t discovered via WebSearch first, unless the user provides the URL directly. Never rely on WebSearch snippets alone — always fetch for facts you’ll cite.