nixos/shared/linked-dotfiles/opencode/skills/research/SKILL.md

name	description
research	Use when conducting deep research across any domain - provides citation patterns, concise synthesis techniques, and cross-domain connection strategies

Research Skill

Quick reference for conducting deep research with proper citations, concise output, and novel insights across any domain.

When to Use This Skill

• Investigating technical topics (APIs, frameworks, algorithms)
• Understanding psychological or human factors
• Exploring creative writing techniques or artistic approaches
• Synthesizing information from multiple disparate sources
• Making connections between different domains
• Gathering authoritative sources for decision-making

When NOT to use:

• Simple factual lookups (use direct search instead)
• Code implementation (use coding agents)
• Quick reference checks (use man pages directly)

Citation Format Reference

Web Sources

<source url="https://example.com/article" title="Article Title">Specific claim or finding from the source</source>

Academic Papers

<source url="https://arxiv.org/abs/2210.03629" title="ReAct: Synergizing Reasoning and Acting">Quantitative finding or key insight</source>

Local Files

<source url="file:///path/to/file.md" title="filename.md">Code snippet or configuration detail</source>

Man Pages

<source url="man://grep" title="grep(1) manual">Command behavior or flag description</source>

Multiple Sources for Same Claim

<source url="https://source1.com" title="First Study">Initial finding</source> corroborated by <source url="https://source2.com" title="Second Study">confirming evidence</source>

Output Style Guide

Good (Concise Paragraphs)	Bad (Bullets/Verbosity)
The STORM paper introduces a multi-agent system for comprehensive research. Finding. This approach yields Wikipedia-quality output.	It is important to note that: - STORM uses agents - Perhaps it works well - Furthermore, one might consider...
Direct statement with citation.	"It seems that this might be useful..."
2-4 sentence paragraphs.	Wall of text or excessive bullets.

ReAct Research Loop

Pattern for iterative research:

1. Thought: "I need to understand X. Sources to check: Y, Z."
2. Action: curl https://docs.example.com/api or man command or rg "pattern"
3. Observation: "Found A, B, C. Still missing D."
4. Thought: "Need to refine search for D."
5. Action: New search with refined query
6. Observation: "Now have complete picture."

Repeat until sufficient evidence gathered.

Making Cross-Domain Connections

Technique: Analogy Mapping

1. Identify core mechanism in source domain
2. Find parallel structure in target domain
3. Map relationships explicitly
4. Test if analogy reveals new insights

Example: ReAct pattern (technical) ↔ Expert problem-solving (psychology)

• Both externalize thinking
• Both enable error detection
• Both reduce cognitive load
• Connection reveals why ReAct works

Technique: Pattern Recognition

Look across sources for:

• Recurring themes or principles
• Shared constraints or trade-offs
• Similar solution approaches
• Common failure modes

Technique: Contrast Analysis

When sources disagree:

• Identify specific points of tension
• Examine underlying assumptions
• Consider context differences
• Synthesize higher-level insight

Research Tool Usage

Web Research

# Fetch documentation
curl -s https://docs.example.com/api | grep "pattern"

# Download paper
wget https://arxiv.org/pdf/2210.03629.pdf

# Search with specific terms
curl -s "https://api.example.com/search?q=term"

Man Pages

# Full manual
man grep

# Search within man page
man grep | grep -A 5 "pattern"

# List all man pages for command
man -k search_term

Code/File Research

# Find implementations
rg "function_name" --type rust

# Search with context
rg "pattern" -A 3 -B 3

# Find files by name
find . -name "*.md" -type f

Common Mistakes

Mistake: Uncited claims

❌ "The ReAct pattern improves performance significantly."
✅ "The ReAct pattern improves performance significantly. <source>specific metric</source>."

Mistake: Verbose hedging

❌ "It seems that perhaps one might consider that this could potentially..."
✅ "This approach increases success rates by 34%."

Mistake: Bullet point overload

❌ Long lists of disconnected bullets
✅ 2-4 sentence paragraphs that flow coherently

Mistake: Missing cross-domain insights

❌ Only technical analysis without broader connections
✅ "This pattern mirrors cognitive psychology research on expert problem-solving..."

Mistake: No verification step

❌ Output without checking citation coverage
✅ Self-check: Every claim cited? Format correct? Under 500 words?

Output Structure Template

[1-2 sentence context establishing the question and why it matters]

[Paragraph 1: First major finding with citations. 2-4 sentences. Focus on one insight.]

[Paragraph 2: Second major finding with citations. Builds on or contrasts with first.]

[Paragraph 3: Cross-domain connection or novel insight. "Interestingly..." or "This pattern mirrors..."]

[Optional Paragraph 4: Practical implications or actionable takeaways if relevant.]

## Sources
1. [Title](URL) - Brief description of what this source provides
2. [Title](URL) - Brief description

Quality Checklist

Before finalizing research output:

• Every significant claim has source citation
• Citations use correct XML format with URL and title
• Writing is direct - no hedging or filler phrases
• Output uses small paragraphs (2-4 sentences), not bullets
• Under 500 words for typical queries
• At least one cross-domain connection identified
• Multiple perspectives considered (technical, human, business, creative)
• Bibliography section lists all sources
• Verification step completed

Domain-Specific Notes

Technical Research

Focus on: Implementation details, performance metrics, trade-offs, compatibility, version-specific behavior

Psychology/Human Factors

Focus on: User studies, cognitive principles, behavioral patterns, accessibility, mental models

Creative Writing

Focus on: Techniques from literature, stylistic approaches, examples from published works, craft advice from authors

Science/Research

Focus on: Peer-reviewed sources, methodology, empirical findings, replication status, statistical significance

Example: Good Research Output

Query: How does the ReAct reasoning pattern work?

Output:

The ReAct pattern addresses a key limitation in language model task-solving by interleaving reasoning traces with tool actions. ReAct agents achieve 34% higher success rates on ALFWorld tasks and 10% improvement on HotpotQA compared to baselines. The core mechanism involves explicit Thought-Action-Observation loops where the model verbalizes its reasoning before taking each action, then updates its understanding based on observations.

This architecture mirrors findings from cognitive psychology on expert problem-solving. Experts externalize their thinking through verbal protocols, which reduces cognitive load by offloading working memory to external representations. By forcing LLMs to "think aloud," ReAct essentially implements this expert strategy in artificial systems.

The practical implication is clear: tasks requiring multi-step reasoning benefit from explicit trace generation. ReAct excels when error recovery matters, since failed actions produce observations that redirect reasoning. For simple, single-step tasks, the overhead isn't justified.

Sources

1. ReAct: Synergizing Reasoning and Acting in Language Models - ICLR 2023 paper introducing the pattern
2. Expert Problem Solving Research - Cognitive science on externalized thinking

---

Note: This skill is designed for use with the research agent. It provides quick reference patterns for citations, concise synthesis, and cross-domain insights.