web-research

---

name: web-research model: sonnet description: Produces a structured Research Summary document (findings, sources, attributions) from targeted web queries using Brave Search and web_fetch. General fact-finding and claim verification. For researching how to use a specific library/API/framework toward an implementation handoff, use web-research-external instead. Use when: 'research this online', 'find current information about X', 'verify this claim', 'gather sources on Y', 'search for what the field says about Z'. category: learn-research

inputs:

• name: query type: string description: The research question or topic to investigate online required: true outputs:
• name: research_summary type: ref format: cas-ref description: Structured Research Summary document with findings, sources, and attributions from web queries

---

Web Research Skill

Version: 1.1 Created: 2026-02-02 Author: Cipher (self-taught) Purpose: Effective web research using Brave Search API and web_fetch for content extraction

---

Overview

This skill encodes best practices for web research — finding, evaluating, and synthesizing information from online sources. It provides patterns for search query formulation, source evaluation, information synthesis, and attribution. Web research is not about collecting links — it's about building understanding from reliable sources.

Output templates for all research modes are in references/research-output-templates.md.

---

Core Tools

Available Tools:

• web_search(query, count, country, search_lang, ui_lang, freshness) — Brave Search API for finding sources
• web_fetch(url, extractMode, maxChars) — Extract readable content from URLs (markdown/text)

Parameters:

• query: Search query string (required)
• count: Number of results (1-10, default: 5)
• country: 2-letter country code for regional results (default: 'US')
• search_lang: ISO language code for search results
• ui_lang: ISO language code for UI elements
• freshness: Time filter for results (pd=past 24h, pw=past week, pm=past month, py=past year)
• url: HTTP or HTTPS URL to fetch
• extractMode: 'markdown' or 'text' (default: 'markdown')
• maxChars: Maximum characters to return (truncates when exceeded)

---

When to Use This Skill

• Investigating a topic that requires current information
• Finding sources for research or specifications
• Verifying claims or facts
• Gathering competitive intelligence or landscape scans
• Understanding a technology, framework, or practice
• Finding documentation or examples

---

Research Workflow

Step 1: Define Your Research Question

Before searching, clarify:

1. What specific question am I trying to answer?
2. What level of detail do I need? (overview vs. deep dive)
3. Is this time-sensitive? (recent events, API versions, industry news)
4. What context do I already have? (avoid re-searching known info)

Examples:

• Bad: "AROMA"
• Good: "AROMA agent collaboration architecture v0.0.1 specification"
• Good: "AI agent memory management best practices 2026"

Step 2: Formulate Search Queries

Pattern: [topic] [context/aspect] [specific keywords] [optional: date filter]

Examples:

Topic	Bad Query	Good Query
AROMA	"AROMA"	"AROMA agent collaboration v0.0.1 specification"
Next.js bugs	"Next.js bugs"	"Next.js 16.1.6 security vulnerabilities CVE"
AI research	"AI research"	"AI agent memory compression techniques 2025"
Excel design	"Excel design"	"Excel data visualization best practices conditional formatting"

Freshness filters:

• Breaking news: freshness="pd" (past 24 hours)
• Recent developments: freshness="pw" (past week)
• Tech releases: freshness="pm" (past month)
• Historical context: freshness="py" (past year) or no filter

Step 3: Search and Select Sources

Execute search:

web_search(
  query="your query here",
  count=5,
  country="US",
  search_lang="en"
)

Evaluate each source:

Criterion	What to Check	Good Signs	Bad Signs
Credibility	Domain reputation	.gov, .edu, established news	Unknown blog, social posts
Relevance	Title and snippet match	Direct answer to question	Tangential content
Freshness	Publication date	Recent (past 1-3 months)	Outdated (>1 year old)
Depth	Content length and detail	500+ words, specific details	200-word overview
Authority	Author or org expertise	Named experts, official docs	Anonymous, generic content

Source selection:

• Prioritize: Official docs, established news, expert publications
• Supplement: Community discussions, GitHub issues, forums
• Verify: Cross-reference claims across 2-3 sources

Step 4: Fetch and Extract Content

For relevant sources:

web_fetch(
  url="https://example.com/page",
  extractMode="markdown",
  maxChars=20000
)

Extraction strategy:

1. Read headers first — Understand structure, main sections
2. Extract key insights — 1-3 sentences per section
3. Note supporting details — Numbers, names, dates, versions
4. Skip filler — Introduction fluff, generic advice
5. Capture sources cited — Link back to original content

Step 5: Synthesize Findings

Goal: Answer the research question, not regurgitate content.

Synthesis template — see references/research-output-templates.md for the full Research Synthesis format.

Synthesis principles:

• Be specific — Avoid "some say," "likely," "possibly"
• Attribute claims — "According to [Source], X is true" not "X is true"
• Note contradictions — "Source A claims X, but Source B says Y"
• Signal uncertainty — "Could not verify" or "Limited evidence available"

---

Research Modes

Mode 1: Verification Research

Use when: Verifying a specific claim, fact, or data point

Process:

1. Formulate specific query: "[claim] verify"
2. Search 3-5 sources
3. Cross-reference across sources
4. Note consensus or conflict

Output: Verification document with Verdict (Confirmed / False / Partially Confirmed / Could Not Verify), Evidence section listing each source's position — see references/research-output-templates.md

---

Mode 2: Deep Dive Research

Use when: Need comprehensive understanding of a complex topic

Process:

1. Start with overview query: "[topic] overview"
2. Identify subtopics from results
3. Query each subtopic: "[topic] [subtopic] details"
4. Fetch and read 2-3 sources per subtopic
5. Synthesize into structured overview

Output: Deep Dive document organized by subtopics with cross-subtopic synthesis — see references/research-output-templates.md

---

Mode 3: Landscape Scan

Use when: Broad pattern recognition across many topics

Process:

1. Formulate 5-10 related queries
2. Execute searches (1-2 sources each)
3. Extract patterns and themes
4. Create opportunity or comparison matrix

Output: Landscape Scan document with themes, key players table, and gaps list — see references/research-output-templates.md

---

Quality Checklist

Before concluding web research, verify:

Source Quality

• At least 2-3 sources consulted (unless verification)
• Sources are credible (official docs, established news, experts)
• Sources are recent (within 1-3 years, unless historical context)
• Conflicting information is noted

Content Quality

• Specific claims are supported by evidence
• Attribution is clear (which source said what)
• Uncertainty is signaled (couldn't verify, limited data)
• Filler is minimized (no generic advice)

Synthesis Quality

• Research question is answered directly
• Key findings are specific and actionable
• Sources are cited with URLs
• Open questions or gaps are noted

---

Common Pitfalls to Avoid

• Single-source confirmation — Finding one source that confirms belief — cross-reference across 2-3 sources
• Over-fetching — Reading 50 pages for one query — focus on 2-5 relevant sources
• Generic queries — "AI tools" returns 10M results — use specific queries with context
• No attribution — "Studies show X" — attribute: "According to [Source], X"
• Outdated data — Using 2019 info for a 2026 decision — use freshness filter
• Content dumping — Copying entire articles — extract key insights (1-3 sentences per section)

---

Integration with Other Skills

Use with:

• research-modes — For structuring deep vs. wide research phases
• specification-writer — When research feeds into spec writing
• seed-extraction — When research reveals reusable patterns
• workspace-navigation — When organizing research findings in shared workspaces

Pattern:

web_search() -> web_fetch() -> Extract insights ->
  research-modes(structure) OR specification-writer(draft) OR seed-extraction(capture)

---

Output

• A Research Summary markdown document answering the research question
• Saved to the project's docs/research/ or scouts/ directory
• Named: [date]_[topic]_web_research.md
• Includes: summary, key findings, supporting details, cited sources with URLs, open questions

---

Examples

Scenario 1: "Find current information about WebSocket performance benchmarks for Go" → Verification/Deep Dive document with 3-5 authoritative sources, specific numbers, attribution per claim, and open questions flagged

Scenario 2: "Search for what the field says about AI agent orchestration patterns" → Landscape Scan document with theme clusters, key tool/framework table, and identified gaps across 8-10 sources

---

Edge Cases

• Paywalled content: Note the source in the synthesis as "paywalled, snippet only" and extract what is available from the search snippet; do not fabricate full content
• Contradictory authoritative sources: Surface the contradiction explicitly; do not pick one silently — present both and note the disagreement
• Topic with no recent sources (>2 years old): Label findings as potentially outdated; recommend confirming with a domain expert or official changelog
• Query returns zero relevant results: Reformulate with 2-3 alternative phrasings before concluding the topic is unresearchable

---

Anti-Patterns

• Fetching the full text of 10+ pages when 2-3 targeted extractions answer the question — token cost with no quality gain
• Summarizing sources in the order they were found rather than organizing findings by theme — produces a list, not a synthesis
• Accepting a source as authoritative because its domain sounds credible without checking the actual author or date
• Using web research for questions that can be answered from the existing codebase or memory — always check local context first

---

Related Skills:

• research-modes — Deep vs. wide research structuring
• specification-writer — Research to spec conversion
• seed-extraction — Pattern extraction from research findings

---

Last Updated: 2026-04-08 Maintained By: Cipher Status: Active