Research Skill

Coordinates parallel research sub-agents to produce thorough, sourced findings on any technical topic.

When to Use

Invoke this skill when you need information that spans multiple sources or domains:

• "What's the best way to handle X in language Y?"
• "How does our codebase currently handle Z?"
• "What do the official docs say about feature W?"
• "Research approaches to solving problem P"

Usage

/skill:research <query>

Or invoke within a task by asking Claude to "research [topic] using the research skill."

How It Works

The skill spins up specialized sub-agents for each research mode relevant to the query:

Sub-Agent: Web Researcher

• Uses WebSearch and WebFetch
• Targets official docs, GitHub, technical blogs, release notes
• Validates source recency and authority
• Returns: findings with URLs and publication dates

Sub-Agent: Codebase Researcher

• Uses Grep, Glob, and Read
• Maps existing implementations, patterns, and usage
• Identifies dependencies, callers, and conventions
• Returns: file paths, line numbers, and code snippets

Sub-Agent: Documentation Researcher

• Fetches API references, RFCs, and specifications
• Extracts relevant signatures, constraints, and examples
• Returns: structured documentation excerpts with links

Output Format

## Research Report: <topic>

### Summary
High-level answer to the research question.

### Findings by Source

#### Web Sources
- [Source Title](url): Key finding

#### Codebase
- `path/to/file.ts:42` — description of what's there

#### Documentation
- [Official Docs](url): Relevant excerpt

### Synthesis
Integrated answer drawing from all sources.

### Sources
Complete list of all URLs and file references used.

### Next Steps
Recommended actions based on findings.

Configuration

Sub-agents run in parallel by default. To run sequentially (e.g., when codebase findings should inform web research), specify in your prompt:

Research X sequentially: first check the codebase, then search the web for best practices.

Notes

• Always cite sources — never present findings without attribution
• Flag when sources conflict and explain the trade-offs
• Mark findings with their date when recency matters
• For codebase research, always verify findings with Read before reporting