Research
Overview
Structured research process that works for coding tasks (evaluating libraries, understanding APIs) and non-coding work (market research, content creation, learning new domains). Produces organized, actionable artifacts.
Process
1. Scope
Define what we're researching and why:
- What question(s) need answering?
- What decisions will this inform?
- What's out of scope?
- What format should the output take?
Confirm scope with user before proceeding.
2. Prior Work
Before gathering new information, check for existing knowledge:
- Search any available knowledge base for prior research, decisions, or patterns on this topic
- Review project docs, CLAUDE.md, and existing codebase patterns
- Summarize what's already known and identify gaps that still need research
- Note any prior work that's outdated or low confidence and may need updating
If nothing relevant exists, proceed to Gather.
3. Gather
Collect information from available sources:
- Existing knowledge: Prior decisions, patterns, and research (from step 2)
- Code: Read relevant files, grep for patterns, check dependencies
- Web: WebSearch for docs, articles, comparisons. WebFetch for specific pages
- Docs: Framework docs via context7 MCP if available
Use parallel subagents for independent research threads.
4. Organize
Structure findings:
- Group by theme, not by source
- Flag contradictions between sources and existing knowledge
- Note confidence level (verified, likely, uncertain)
- Separate facts from opinions
5. Synthesize
Draw conclusions:
- Answer the original questions directly
- Provide recommendation with reasoning
- List trade-offs explicitly
- Note what remains unknown
6. Artifact
Deliver in the agreed format:
- Decision: Recommendation + alternatives + trade-offs table
- Comparison: Feature matrix with weighted criteria
- Summary: Key findings + action items
- Brief: Background + analysis + recommendation (for sharing with others)
7. Persist
After the user accepts findings, save them for future retrieval:
- If a knowledge base is available, deposit the findings with appropriate metadata (type, tags, confidence, related notes)
- If research supersedes prior work, mark the old artifact as outdated
- Skip for trivial lookups or if the user declines
Quick Reference
| Research Type | Key Sources | Typical Output |
|---|---|---|
| Library eval | npm/pypi, GitHub stars/issues, docs | Comparison matrix |
| Bug investigation | Code, logs, issue trackers | Root cause + fix options |
| Architecture | Existing decisions, code, web patterns | Decision document |
| Content/topic | Web search, articles, papers | Structured summary |
| API integration | API docs, examples, SDKs | Integration guide |
Common Mistakes
- Starting to gather before defining scope (wastes time on irrelevant info)
- Skipping prior work check (re-researching what's already known)
- Presenting raw findings without synthesis (user wants answers, not data dumps)
- Not confirming scope with user (researching the wrong thing)
- Single-source conclusions (verify across multiple sources)
- Forgetting to persist findings (research evaporates after the session)