research

name	research
description	Deep research methodology for sub-agents and spawn tasks. Use when spawning research tasks, investigating claims, comparing products/services, or gathering comprehensive information on any topic. Ensures source diversity, cross-referencing, bias detection, and confidence assessment. Automatically applied to research-type spawns.

Research Skill

Orchestration Pattern

Model allocation (see workspace SKILLS.md for actual model assignments):

• Research main agent: Deep analysis model — thorough, accurate
• Research sub-agents: Fast parallel model — cheap, high throughput
• Global fallback: If main model fails, automatically use default model

Model Configuration:

• Main agent (researcher): Uses the deep analysis model for synthesis
• Sub-agents (collectors): Use the fast model for parallel source gathering
• This overrides global defaults — research tasks always use this allocation

Why this pairing:

• Deep analysis model: Slow but thorough, checks local files + web, high accuracy for synthesis
• Fast parallel model: Cheap and fast, perfect for gathering raw sources in parallel

When to parallelize:

• Spawn separate sub-agents for unrelated topics within a research task
• Keep related aspects in same sub-agent to preserve context
• Each collector returns structured findings; main agent synthesizes

Example flow:

1. Main agent receives research task
2. Identifies unrelated aspects (e.g., "pricing" vs "technical specs")
3. Spawns fast sub-agents in parallel (one per topic) for source gathering
4. Each returns findings in standard format
5. Main agent synthesizes, cross-references, assigns confidence

---

Methodology

Follow these steps for every research task:

1. Query Diversity & Web Search Tool

Use the web search tool configured in workspace SKILLS.md — no rate limits, can spam freely.

• Try multiple phrasings of the search query (synonyms, different angles)
• Spawn fast sub-agents to run searches in parallel
• Include platform-specific searches when applicable:
  • site:reddit.com for community opinions and real experiences
  • site:github.com for technical projects, issues, discussions
  • Official product/company sites for authoritative info

2. Source Volume

• Fetch at least 10 sources before drawing conclusions
• Include date filters: 2025 and 2026 to ensure recency
• Flag any source older than 6 months as potentially outdated

3. Source Classification

For each source, identify:

• Primary: Official docs, announcements, product pages, API references
• Secondary: Blogs, articles, tutorials, reviews
• Community: Reddit, HN, forums, GitHub discussions

Prioritize primary sources for factual claims.

4. Bias Detection

Flag sources exhibiting:

• PR/marketing speech (vague superlatives, no specifics)
• Point of sale (affiliate links, sponsored content)
• Clickbait (sensational headlines, thin content)
• Outdated info presented as current

Note bias in findings; don't discard biased sources entirely—they can still contain useful data.

5. Cross-Referencing

• Every factual claim must appear in 2+ independent sources
• If only one source makes a claim, mark as "unverified"
• Conflict detection: Explicitly note when sources contradict each other

6. Knowledge Gaps

Acknowledge what you couldn't find or verify:

• "No official source confirms X"
• "Pricing info only found on third-party aggregators"
• "No data available after [date]"

7. Confidence Assessment

Rate each major finding:

• High: 3+ primary sources agree, recent, no conflicts
• Medium: 2+ sources agree, or primary + secondary alignment
• Low: Single source, outdated, or conflicting information
• Unverified: Claim found but not corroborated

---

Output Format

Structure all research findings as:

## Executive Summary
[2-3 sentence answer to the research question]

## Key Findings
- Finding 1 [Confidence: High/Medium/Low]
- Finding 2 [Confidence: High/Medium/Low]
...

## Conflicts & Contradictions
[Note any sources that disagree and what they disagree on]

## Knowledge Gaps
[What couldn't be verified or found]

## Sources
1. [Title](URL) - Primary/Secondary/Community - [Date if known]
2. ...

---

Quick Reference

Step	Requirement
Queries	Multiple phrasings, include Reddit/GitHub
Sources	Minimum 10 before conclusions
Dates	Filter for 2025-2026, flag >6mo old
Cross-ref	2+ sources per claim
Bias	Flag PR, affiliate, clickbait
Conflicts	Explicitly note contradictions
Gaps	State what's unknown
Confidence	High/Medium/Low/Unverified per finding