question-refiner

yeheng 的头像作者 yeheng

将原始研究问题细化为结构化的深度研究任务。通过提问澄清需求,生成符合 OpenAI/Google Deep Research 标准的结构化提示词。当用户提出研究问题、需要帮助定义研究范围、或想要生成结构化研究提示词时使用此技能。

📖 中文导读:将原始研究问题细化为结构化的深度研究任务。通过提问澄清需求,生成符合 OpenAI/Google Deep Research 标准的结构化提示词。当用户提出研究问题、需要帮助定义研究范围、或想要生成结构化研究提示词时使用此技能。 以下正文为该技能的原始 SKILL.md 内容(英文),上方安装命令可直接使用。

name: question-refiner description: Transform raw research questions into structured, validated research prompts with automatic research type detection and output format validation. Ensures prompts are ready for research-executor with comprehensive quality checks.

Question Refiner

Overview

Transform vague research questions into structured, actionable research prompts through strategic clarifying questions with automatic research type detection and quality validation.

When to Use

  • User provides a raw, unstructured research question
  • Research scope is unclear or too broad
  • Need validated structured prompt for research-executor
  • Want to ensure prompt meets quality standards (≥8.0)

Core Approach

Progressive Questioning (2 rounds max):

  1. Round 1 (3 questions): Topic focus, output format, audience
  2. Round 2 (conditional): Scope, sources, special requirements
  3. Auto-detect research type → Select template → Generate & validate

Research Type Detection

Type Indicators Example
Exploratory "what is", "overview", "landscape" "What is the AI market like?"
Comparative "vs", "compare", "difference" "Compare GPT-4 vs Claude"
Problem-Solving "how to", "solve", "fix" "How to improve API performance"
Forecasting "future", "trend", "prediction" "Future of quantum computing"
Deep Dive "technical", "architecture" "How does BERT work internally"
Market Analysis "market", "industry", "competition" "AI chip market analysis"

Output Structure

### RESEARCH TYPE
[auto-detected type]

### TASK
[Clear, specific research objective]

### CONTEXT/BACKGROUND
[Why this matters, who will use it]

### SPECIFIC QUESTIONS
1-7 concrete sub-questions

### KEYWORDS
[Search terms ≥5]

### CONSTRAINTS
- Timeframe: [e.g., 2020-present]
- Geography: [e.g., global]
- Source types: [academic, industry, news]

### OUTPUT FORMAT
- Type: [comprehensive_report|executive_summary|comparison_table]
- Citation style: [inline-with-url|footnotes]

### QUALITY SCORE
[0-10, must be ≥8.0]

Quality Validation

Component Weight Criteria
Completeness 30% All required fields present
Specificity 30% Questions are specific, not vague
Keyword Richness 20% ≥5 search terms with synonyms
Constraint Clarity 20% Clear, realistic constraints

Process: Generate → Validate → If score < 8.0: Refine (max 2 attempts)

Token Optimization

📋 Reference: .claude/shared/constants/token_optimization.md

Context Budget: 10k tokens max

Error Handling

📋 Reference: .claude/shared/constants/error_codes.md

  • E001: Insufficient context → Ask clarifying questions
  • E003: Validation failed → Refine and retry
  • E004: Quality < 8.0 after retries → Request manual review

See also: Skill Base Template

Examples

See examples.md for detailed interaction patterns.

Detailed Instructions

See instructions.md for complete questioning strategy.