---

name: web-research description: Use when any agent needs deep web research on a topic — gathering, verifying, and structuring information from multiple sources. Covers search strategy, source evaluation, parallel research agents, confidence levels, gap flagging, and structured output. Invoke this instead of ad-hoc searching.

Web Research

Overview

Structured deep research that produces verified, organised findings with confidence levels and source evaluation. Any agent can invoke this skill when they need information — it replaces ad-hoc searching with a systematic methodology. Output is saved to research/[topic]/ for reuse.

When to Use

• Agent needs factual data to make a decision
• Challenger needs to verify claims
• Forge needs domain context before design
• research-for-skills needs domain knowledge before writing a skill
• Any task where "I think" should be replaced with "the evidence shows"

Internet Reachability Check

Before running web searches, check if internet is available via the knowledge-grounding manifest:

INTERNET=$(jq -r '.internet_reachable' ~/.claude/state/sources.json 2>/dev/null || echo "unknown")

If internet_reachable is false:

• Do NOT attempt web searches (they will hang or fail)
• Suggest local sources instead: wikis, project docs, configured doc_paths
• Read active sources from ~/.claude/state/sources.json to suggest alternatives

If the manifest does not exist, run bash ~/.claude/skills/knowledge-grounding/scripts/discover.sh discover --silent first.

Antigravity (agy) Cross-Verification

When available, the Antigravity CLI (agy) provides an additional research source from a different model family. Invoke it headless with stdin closed (mandatory — see the antigravity-cli skill):

timeout 600 agy -p "Verify: <claim>. Cite sources." < /dev/null

Use it to:

• Cross-verify findings from WebSearch with a different model/search backend
• Get second-model answers for freshness-sensitive topics (latest versions, recent changes)
• Run parallel verification: WebSearch + agy + Codex (/codex:rescue) for triangulation
• Synthesize large research outputs that exceed Claude's working context (agy handles large-context analysis well)

(See codex-orchestration / cross-cli-deliberation for full delegation patterns.)

The Research Process

1. SCOPE    → Define what we need to know (sub-questions)
2. SEARCH   → Parallel agents search different angles
3. EVALUATE → Apply source evaluation to every finding
4. VERIFY   → Triangulate across 3+ independent sources
5. FLAG     → Mark what was NOT found (gaps)
6. STRUCTURE → Organise with confidence levels
7. SAVE     → Store in research/[topic]/

Step 1: Scope — Decompose the Question

Break the research topic into 4-5 searchable sub-questions using MECE (Mutually Exclusive, Collectively Exhaustive):

Topic: "Is Redis suitable for our session storage?"

Sub-questions:
1. What are Redis's session storage capabilities and limits?
2. What are the alternatives (Memcached, DB, JWT) and trade-offs?
3. What are Redis failure modes and data loss risks?
4. What's the operational cost (hosting, monitoring, maintenance)?
5. What do practitioners report from production use?

Rule: If you can't break it into sub-questions, the topic is too vague. Refine it first.

Research Levels

Choose the level BEFORE starting. The calling agent or user specifies the level. When not specified, use the decision guide below.

SHORT — Quick Verification (seconds to 1 minute)

Use when: Verifying a single claim, checking if a library exists, confirming a version number, getting one specific fact.

Aspect	Detail
Agents	0 (inline WebSearch by the requesting agent)
Searches	1-3 targeted queries
Source eval	Quick check — is source official/reputable?
Triangulation	Not required for factual lookups (docs, versions)
Output	Inline answer with source link — no research folder
Gap flagging	"Couldn't verify" if not found in 3 searches

Examples: "Does ibm_db support Python 3.14?", "What's the current Flask version?", "Is FAQ schema still generating rich results?"

MEDIUM — Focused Research (2-5 minutes)

Use when: Understanding a specific topic, comparing 2-3 options, getting current best practices for a focused area, verifying multiple related claims.

Aspect	Detail
Agents	1-2 parallel research agents
Searches	5-15 per agent
Source eval	SIFT method on key findings
Triangulation	Required for claims that inform decisions
Output	Brief findings doc — save to research/ if reusable
Gap flagging	Required
Confidence levels	Required on key findings

Examples: "Best deployment pattern for Flask + Gunicorn in 2026", "Redis vs Memcached for session storage", "Current state of Google FAQ schema support"

LONG — Deep Research (5-15 minutes)

Use when: Creating a new skill, understanding an entire domain, multi-faceted topic with many subtopics, need comprehensive coverage with challenger review.

Aspect	Detail
Agents	3-5 parallel research agents + challenger
Searches	10-30 per agent
Source eval	Full SIFT + trust hierarchy on all findings
Triangulation	Required on all key claims (3+ sources)
Output	Full research folder: INDEX.md + per-subtopic files + sources
Gap flagging	Required — explicit "NOT FOUND" table
Confidence levels	Required on every finding
Challenger	Reviews all findings for contradictions, bias, gaps

Examples: "Full SEO landscape 2026", "Python parallelism patterns", "WooCommerce security best practices", "Docker admin patterns"

Decision Guide — How to Choose Level

Is this a single fact or version check?
  → SHORT

Is this a focused question with 1-2 sub-questions?
  → MEDIUM

Does it have 3+ sub-questions, or will it become a skill?
  → LONG

Still unsure? Ask:
  "How many sub-questions does this break into?"
    1-2  → MEDIUM
    3+   → LONG
    0    → SHORT (it's just a fact check)

Callers can override: An agent requesting research can specify the level. If not specified, use the decision guide. When in doubt, start MEDIUM — escalate to LONG only if findings are insufficient.

Step 2: Search — Execute at Chosen Level

SHORT: Inline Search

No agents spawned. The requesting agent runs 1-3 WebSearch queries directly and returns the answer with source.

MEDIUM: Focused Agents

Spawn 1-2 research agents, each covering a sub-question:

"Research [SUB-QUESTION]. Find:
- Current best practices (2024-2026)
- Data points with specific numbers and sources
- Official documentation references
- Contradicting viewpoints if they exist

For every claim, note the source URL and date.
Flag anything you searched for but couldn't find.
DO NOT make up data. If uncertain, say so."

LONG: Full Research Team

Spawn 3-5 parallel agents + 1 challenger:

Research agents (parallel):
"Research [SUB-QUESTION] deeply. Find:
- Current best practices with specific data/tools/versions
- What's changed since 2023-era advice
- Anti-patterns — common mistakes and outdated advice
- Benchmarks/data points to include
- Practitioner experience (Reddit, HN, Stack Overflow)
- Actionable rules an AI agent should follow

Use WebSearch extensively. Be specific — no generic advice.
For every claim, note source URL and date.
Flag anything searched for but not found."

Challenger agent (after research completes):
"Review all findings for contradictions, vendor bias, outdated
claims, and gaps. Verify key statistics. Flag questionable data."

Step 3: Evaluate — Source Assessment

Apply the SIFT method (Stop, Investigate, Find better, Trace claims) to every source:

Source Trust Hierarchy

Tier	Source Type	Trust	Action
1	Official documentation (Google, MDN, RFCs, specs)	HIGH	Accept, check date
2	Peer-reviewed / transparent methodology studies	HIGH	Check methods, sample size
3	Independent research with data (Ahrefs, Semrush with methodology shown)	MEDIUM-HIGH	Note potential vendor bias
4	Reputable industry publications (Search Engine Journal, Smashing Magazine)	MEDIUM	Cross-reference claims
5	Blog posts, tutorials, forum answers	MEDIUM-LOW	Verify claims independently
6	Vendor marketing, press releases, case studies	LOW	Assume biased — seek independent confirmation
7	Unsourced claims, "everyone knows", AI-generated stats	REJECT	Demand evidence or discard

Red Flags in Sources

• No date published (reject for any fast-moving topic)
• Statistics without named study/methodology
• Same stat cited across 10 blogs but no original source (circular citation / citogenesis)
• "Studies show" without naming the study
• Vendor's own case study proving their product works
• Numbers that are too precise for the claim ("exactly 94% accuracy")

Step 4: Verify — Triangulation

Every key finding must be triangulated across 3+ independent sources.

Triangulation	When
Same fact, 3+ independent sources	Standard verification — fact is reliable
2 sources agree, 1 contradicts	Note the contradiction, investigate why
Only 1 source	Mark as "single source — unverified"
Sources disagree fundamentally	Mark as "contradictory" — present all sides
No sources found	Mark as "NOT FOUND" — this is valuable information

Deliberately search for contradicting evidence: After finding support for a claim, search for "[claim] wrong", "[claim] criticism", "[claim] debunked".

Step 5: Flag — Information Gaps

Explicitly state what you searched for but could NOT find. This is as valuable as what you did find.

### Information Gaps

| Searched For | Result |
|-------------|--------|
| Redis session storage benchmark at 100K concurrent users | NOT FOUND — no public benchmarks at this scale |
| Comparison of Redis vs Valkey for sessions (post-fork) | INSUFFICIENT — only vendor blog posts, no independent comparison |
| GDPR implications of Redis session data in EU | PARTIAL — found general guidance but no Redis-specific legal opinion |

Step 6: Structure — Confidence Levels

Every finding gets a confidence level:

Level	Meaning	Criteria
VERIFIED	Multiple independent sources agree, official docs confirm	3+ sources, at least 1 tier-1
LIKELY	Good evidence but not fully triangulated	2 sources or 1 tier-1 source
UNCERTAIN	Limited evidence, single source, or dated	1 source, or all sources >12 months old
CONTRADICTORY	Sources disagree	Present all sides with sources
NOT FOUND	Searched but no evidence found	Document what was searched

Step 7: Save — Research Output

Save all findings to research/[topic]/:

research/
  [topic]/
    INDEX.md              # Summary of findings + confidence levels
    R1-[subtopic].md      # Detailed findings per sub-question
    R2-[subtopic].md
    sources.md            # Annotated source list

INDEX.md Template

# Research: [Topic]

**Date:** [YYYY-MM-DD]
**Requested by:** [Agent/User]
**Sub-questions:** [List]

## Key Findings

| # | Finding | Confidence | Source |
|---|---------|-----------|--------|
| 1 | [Finding] | VERIFIED | [Source + date] |
| 2 | [Finding] | LIKELY | [Source + date] |
| 3 | [Finding] | CONTRADICTORY | [Source A] vs [Source B] |

## Information Gaps

| Searched For | Result |
|-------------|--------|
| [Query] | NOT FOUND / INSUFFICIENT |

## Contradictions Found

| Claim | Source A Says | Source B Says |
|-------|-------------|---------------|
| [Topic] | [Position + source] | [Counter-position + source] |

## Detailed Findings

See individual research files (R1, R2, etc.) for full evidence.

Integration with Other Skills

Skill	How It Uses web-research
forge	Domain context before design phase
challenger	Verify claims in proposals (self-research capability)
research-for-skills	Gather domain knowledge before writing a skill
seo-* skills	Verify SEO data, check current algorithm status
development-lifecycle	Research phase before implementation

Search Technique Quick Reference

Google Advanced Operators

Operator	Purpose	Example
site:	Search within one domain	site:developers.google.com structured data
"exact phrase"	Match exact words	"Knowledge Graph" entities 2025
before: / after:	Date restrict	Redis session storage after:2025-01-01
filetype:	Specific file types	filetype:pdf OWASP top 10 2025
intitle:	Words in page title	intitle:benchmark Redis vs Memcached
-term	Exclude results	Python parallelism -tutorial -beginner
OR	Either term	"Redis sessions" OR "Valkey sessions"

Platform-Specific Searches

Need	Search Where
Official documentation	site:developers.google.com, site:docs.python.org
Academic/research	Google Scholar, site:arxiv.org
Practitioner experience	site:reddit.com, site:news.ycombinator.com
Code examples	site:github.com, site:stackoverflow.com
Security advisories	site:nvd.nist.gov, site:cve.mitre.org
Current news/changes	Google News, after: date operator

Anti-Patterns

Don't	Why	Do Instead
Accept first search result as truth	First-result bias	Check 3+ sources minimum
Only search for confirming evidence	Confirmation bias	Deliberately search for counter-evidence
Use undated sources for tech topics	Information decays fast	Require dates on all tech/SEO sources
Cite vendor case studies as independent proof	Vendor bias	Flag as "vendor source" and seek independent confirmation
Say "no information exists" after 1 search	Insufficient effort	Try 3+ query variations before declaring NOT FOUND
Present all findings at same confidence	Misleading	Use confidence levels on every finding
Skip gap flagging	Hides ignorance	Explicitly state what wasn't found
Re-research what's already in research/ folder	Waste	Check existing research first