research-methodology（研究方法）

Research Methodology for Documentation

This skill provides systematic approach to researching technical documentation using WebSearch and WebFetch tools.

Core Principles

1. Initialize first - Ensure project knowledge base skill exists
2. Validate before research - Ensure request is specific enough
3. Check local first - Look in .claude/skills/project-knowledge-base/references/ before searching
4. Official sources priority - Start with official docs
5. Filter aggressively - Extract only what's relevant to context
6. Save for reuse - Document findings in standard format

Request Validation

A valid research request must contain three elements:

If any element is missing, return validation error and request clarification.

Search Strategy

Query Formulation

Build queries progressively:

Level 1 (Official): {technology} official documentation {topic}
Level 2 (Tutorial): {technology} {topic} tutorial example
Level 3 (Problem): {technology} {topic} {error-message} solution

Source Hierarchy

Prioritize sources in this order:

1. Official documentation (always check first)

- react.dev, docs.python.org, effect.website - GitHub official repos and examples

1. Trusted secondary sources

- MDN Web Docs (web technologies) - DigitalOcean Community tutorials - Dev.to (high-quality articles only) - Stack Overflow (accepted answers)

1. Avoid

- SEO-optimized content farms - Outdated tutorials (check dates) - AI-generated summaries - Forums without accepted solutions

WebSearch Patterns

Reference references/query-patterns.md for specific query templates per technology domain.

Filtering Results

Relevance Criteria

Include information that:

• Directly addresses the stated context
• Provides actionable code examples
• Explains common pitfalls for the use case
• Is current (matches stated version or latest)

Exclude information that:

• Is tangentially related
• Covers advanced edge cases not needed
• Is deprecated or version-mismatched
• Duplicates what's already found

Extraction Process

1. Scan search results for relevance
2. Open 2-3 most promising sources
3. Extract specific sections, not entire pages
4. Verify code examples are complete
5. Note version compatibility

Document Format

Save all knowledge files to .claude/skills/project-knowledge-base/references/ using the template in references/document-template.md.

File Naming

Format: {technology}-{topic}.md

Examples:

• react-useeffect-cleanup.md
• effect-pipe-operator.md
• prisma-relations.md
• nextauth-jwt-session.md

Rules:

• All lowercase
• Hyphens between words
• Technology first, then topic
• No version numbers in filename

Frontmatter Structure

Required fields in YAML frontmatter:

• topic: Descriptive title
• technology: Library/framework name
• version: Version researched (or "latest")
• sources: List of URLs used
• created: Date in YYYY-MM-DD format
• context: Original problem that triggered research

Progressive Disclosure

Threshold: 500 lines

When a reference file exceeds 500 lines, split into a tree structure:

references/tanstack-router.md  (>500 lines)
↓ split to
references/tanstack-router/
├── _index.md        # Overview + TOC linking to sub-files
├── route-guards.md
├── data-loading.md
└── navigation.md

When to Split

• Single reference file exceeds 500 lines
• Topic has clearly distinct sub-topics
• Different aspects serve different use cases

Split Structure

1. _index.md: Overview, quick reference, TOC with links
2. Sub-files: One file per major sub-topic
3. Cross-references: Link between related sub-files

SKILL.md Update

After splitting, update SKILL.md index to reference _index.md:

- [tanstack-router](references/tanstack-router/_index.md) - TanStack Router comprehensive guide

Quality Checklist

Before saving knowledge document, verify:

• Project knowledge base skill initialized
• Request was properly validated
• Existing knowledge was checked first
• Official sources were consulted
• Content is specific to stated context
• Code examples are complete and tested
• Sources are cited
• File follows naming convention
• Frontmatter is complete
• SKILL.md index updated
• Progressive disclosure applied if >500 lines

Additional Resources

Reference Files

• references/query-patterns.md - Technology-specific search query templates
• references/document-template.md - Complete knowledge document template

Implementation Notes

This methodology is designed for Haiku model execution. Instructions are explicit and procedural to ensure consistent results across model capabilities.