跳至正文
OpenAgentSkill
浏览已开放 · 目录记录日期 2026-11-21 · 自动更新和账户服务暂时关闭。
搜索与研究

research

Structured web research with cited sources and confidence levels; runs in a forked subagent, quick (3-5 sources) or deep (8-10 sources, saved to memory/research/). Use when user says "research X", "look into X", "find information about X", "investigate X", "what's the state of X". Do NOT use for questions answerable from the codebase or memory files, for single-fact lookups (just answer, or fetch the one URL), or when the user already gave you the source to read.

onurpolat05 · 已有静态检查记录

安装与使用

运行命令前先阅读来源与依赖说明。复制命令不会自动安装或执行。

npx skills add onurpolat05/ALBA --skill research

技能说明


name: research description: Structured web research with cited sources and confidence levels; runs in a forked subagent, quick (3-5 sources) or deep (8-10 sources, saved to memory/research/). Use when user says "research X", "look into X", "find information about X", "investigate X", "what's the state of X". Do NOT use for questions answerable from the codebase or memory files, for single-fact lookups (just answer, or fetch the one URL), or when the user already gave you the source to read. context: fork agent: general-purpose background: false effort: medium argument-hint: [deep] allowed-tools: [Read, Write, Glob, Grep, WebSearch, WebFetch, mcp__exa__web_search_exa, mcp__firecrawl__firecrawl_search, mcp__firecrawl__firecrawl_scrape]

/research - Web Research

Perform structured web research on any topic. Runs as a subagent (fork) to keep main context clean.

Why background: false

context: fork alone is not enough. Since Claude Code v2.1.218, forked skills default to background: true — they run detached and the result arrives as a later notification instead of in the turn that invoked the skill. Research is a request for an answer now, so this skill pins background: false: the fork still keeps its search noise out of the main context, but the conversation waits for the report and hands it back inline.

If you copy this skill into a workflow where you genuinely want fire-and-forget research (e.g. kicked off by a scheduled task), drop the background: false line.

Requires Claude Code v2.1.218 or later. On older versions the field is ignored and the skill runs inline-blocking as it always did.

Input

/research [topic]
/research [topic] deep

$ARGUMENTS carries the whole invocation; the trailing word deep selects deep mode.

  • Default (quick): 3-5 sources, summary format
  • Deep: 8-10 sources, full analysis with confidence ratings

Process

1. Search Strategy

Use available tools with graceful fallback:

PriorityToolBest For
1stmcp__exa__web_search_exaSemantic search, finding relevant articles
2ndmcp__firecrawl__firecrawl_searchSearch + scrape combined
3rdWebSearch (built-in)Basic web search
4thWebFetch (built-in)Fetch specific URLs

Fallback rule: If an MCP tool is unavailable, skip it and use the next available. WebSearch + WebFetch are always available as baseline.

The Exa and Firecrawl entries in allowed-tools are there so the chain works for people who do have those MCP servers connected. MCP servers are optional in ALBA — if you never install them, those entries simply never match anything and the skill runs on the built-in tools.

2. Gather & Validate

  • Search with 2-3 different query variations
  • Cross-reference findings across sources
  • Note conflicting information explicitly
  • Prefer recent sources (< 1 year old)

3. Synthesize

Don't just concatenate sources. Analyze, compare, and draw conclusions.

Output Format

Quick Research

## Research: [Topic]

### Summary
[2-3 sentences - the key takeaway]

### Key Findings
- [Finding 1] (Source: [name])
- [Finding 2] (Source: [name])
- [Finding 3] (Source: [name])

### Sources
1. [Title](url) - [one-line relevance note]

Deep Research

## Research: [Topic]

### Executive Summary
[2-3 sentences overview]

### Key Findings
- [Finding] *(Confidence: High - 3+ sources confirm)*
- [Finding] *(Confidence: Medium - 1-2 sources)*
- [Finding] *(Confidence: Low - single source, unverified)*

### Detailed Analysis

#### [Theme 1]
[Analysis with source references]

#### [Theme 2]
[Analysis with source references]

### Confidence Assessment
- **Overall:** [High/Medium/Low]
- **Gaps:** [What couldn't be found or verified]
- **Conflicts:** [Any contradictions between sources]

### Sources
1. [Title](url) - [relevance + reliability note]
2. ...

4. Save (Deep only)

For deep research, save output to memory/research/[topic-slug].md. Create memory/research/ directory if it doesn't exist.

Rules

  • Always cite sources with links
  • Never fabricate or assume information
  • State confidence levels honestly
  • If a topic has no good sources, say so clearly
  • Keep quick research under 30 lines
  • Keep deep research under 100 lines
  • Use the user's language for output (match CLAUDE.md language setting)