Research-First 调研工作流引擎

QoderWork 适配版说明：本 Skill 针对 QoderWork 单线程环境调整。所有调研任务在当前对话中串行执行，通过阶段性汇报模拟"子代理"效果。

When to Use

Use this skill when the user needs:

• Technology selection and comparison
• Solution evaluation and feasibility analysis
• Finding existing tools or libraries
• Best practices research
• Competitive analysis
• Industry trends investigation

Research Levels

L1 - Quick Check

• Condition: Confirm usage, check configuration, find files, ≤3 tool calls
• Execution: Direct answer, no report
• Examples: "How to use X", "What's the config for Y"

L2 - Standard Research

• Condition: Technology selection, solution comparison, GitHub search, 5-15 tool calls
• Execution: Multi-phase execution with progress updates
• Examples: "Compare React vs Vue", "Find best PDF library"

L3 - Deep Research

• Condition: Multi-solution comparison, cross-domain, network + code analysis, >15 tool calls
• Execution: Structured multi-phase with intermediate reports
• Examples: "Evaluate microservices architecture", "Research AI deployment options"

Workflow Selection

Execution Steps

Step 1: Deduplication Check (L2/L3 only)

Before starting L2/L3 research:

1. Read project docs/research/INDEX.md if exists
2. Read global ~/.qoderwork/memory-bank/research/INDEX.md
3. Grep keywords to find existing research
4. If found, ask user: Reuse / Refresh / Ignore

Step 2: Level Determination

Based on task complexity:

• User says "check" / "confirm" → L1
• Involves "selection" / "comparison" / "GitHub" → L2
• Involves "unknown domain" / "multi-dimension" / "high-risk" → L3
• Default to L2 if uncertain

Step 3: Scenario Identification

Select W1-W6 workflow based on user intent (see table above).

Step 4: Requirements Clarification (if needed)

If uncertainties exist, use AskUserQuestion (max 2-3 questions). Skip for L1 and when requirements are clear.

Step 5: Execution Declaration

Before execution, show configuration:

[MODE: RESEARCH] Research Initiated
- Level: L{1|2|3}
- Workflow: W{N} {Workflow Name}
- Validation Depth: {🔴|🟡|🟢}
- Timeliness: {🔥|📅|📚}
- Tools: {List of main tools to use}
- Execution: Single-thread with progress updates

Step 6: Execute (QoderWork Single-Thread Mode)

#### L1: Quick Check

• Use Read/Grep/Glob → Direct answer
• No report, no PRD

#### L2: Multi-Phase Research

Phase 1: Code Analysis

• Read existing code if applicable
• Identify patterns and conventions
• Report: "Phase 1 complete: Found X relevant files..."

Phase 2: External Research

• WebSearch/WebFetch for documentation
• Bash(gh search repos) for GitHub search
• Report: "Phase 2 complete: Found Y potential solutions..."

Phase 3: Analysis & Report

• Analyze findings
• Write report to ~/.qoderwork/memory-bank/research/{topic}-{YYYYMMDD}.md
• Return summary to user

#### L3: Structured Deep Research

For complex multi-domain research, use TodoWrite to track progress:

TodoWrite({
  todos: [
    {content: "Domain A: Solution discovery", status: "in_progress"},
    {content: "Domain B: Technical evaluation", status: "pending"},
    {content: "Domain C: Best practices", status: "pending"},
    {content: "Synthesis & report", status: "pending"}
  ]
})

Execute each domain sequentially with intermediate updates.

Step 7: Persist Results

Confirm:

• Report saved to correct path
• INDEX.md updated
• Storage level correct (project vs global)

Step 8: Output + Handoff

#### L1 Output

## Research Summary
### Findings
- [Key finding 1-3]
### Recommendations
[Brief recommendation]

#### L2/L3 Output

## Research Summary

**Report Location**: `{file path}`

### Key Findings (≤10 points)
1. [Finding 1]
2. [Finding 2]
...

### Reuse Analysis
| Solution/Project | Strategy | Reason |
|------------------|----------|--------|
| {Name} | 🟢 Use Directly / 🟡 Adapt / 🔵 Reference / ⚪ Build Own | {Brief} |

### Recommended Solution
[Summary]

### PRD Framework Draft
**Recommended**: [From research]
**Core Features**: [From key findings]
**Technical Dependencies**: [From reuse analysis]
**Risk Items**: [From validation]
**To Clarify**: [From uncertainties]

### Next Steps
[Clear recommendation for next action]

Tool Priority

1. GitHub: Use gh search repos, gh repo view commands
2. Documentation: WebFetch for official docs
3. Search: WebSearch (general), add site: for specific sources
4. Academic: WebSearch with "site:arxiv.org"
5. Chinese: WebSearch with 知乎/掘金/V2EX

Quality Requirements

• Cross-validate key conclusions with multiple sources
• Distinguish "confirmed facts" from "speculation"
• Include source URL for each key claim
• Dig deeper than first-page results
• Update user on progress for L2/L3 (don't go silent)

Research→Action Handoff

• After research, give clear "next step recommendations"
• If coding needed → suggest switching to development mode
• If decision needed → present "options + recommendation" format

QoderWork vs Claude Code Differences