| name | research |
| description | Researches technical questions with authoritative sources and saves work-local or reusable findings. Use for docs, comparisons, APIs, and architecture references. Triggers on: research, investigate, compare, gather references. |
Research Skill
Research a technical question, synthesize reliable evidence, and save findings either work-locally or as reusable notes in .agents/research/.
Workflow
0. Check Existing Context
Before doing new research:
- If the work belongs to an active work item, read
.agents/work/<category>/<work-slug>/index.md and prefer work-local research.md or research/ for task-specific findings.
- Search
.agents/research/ for an existing reusable topic note.
- Check repo docs, attached files, and skill
references/ first.
- If an existing note already answers the question, update it only when stale or incomplete.
1. Define the Research Brief
Capture these before researching:
- Question: What exact question must be answered?
- Depth: Default to a targeted answer unless a deep-dive is explicitly requested.
- Focus: Implementation details, comparisons, best practices, or conceptual overview.
If the request is ambiguous but actionable, state assumptions and proceed. Ask follow-up questions only when ambiguity would materially change the recommendation.
2. Gather Evidence
Use the smallest toolset that can answer the question.
| Tool | Use for |
|---|
| Web search | Find authoritative sources when local docs are insufficient |
| Web page reading | Extract relevant details from specific URLs |
| Oracle or another reasoning model | Synthesize trade-offs after evidence is gathered |
| Repository search | Inspect local implementations and patterns |
| GitHub tooling | Inspect issues, PRs, releases, or external repository examples |
Guidelines:
- Prefer local docs and existing research before external search.
- Prefer official docs, specs, release notes, and maintainer-authored sources.
- Use community posts only to fill gaps in official guidance.
- Target 2-5 high-quality sources instead of broad source dumps.
- Use deeper reasoning after collecting evidence, not as a substitute for evidence.
- If the user asks for latest or recent information, force a fresh fetch when your tools support it.
3. Synthesize Recommendation
Produce:
- The best current recommendation.
- Key trade-offs.
- Confidence level and uncertainty.
- Unresolved risks or follow-up validations.
4. Save or Update Findings
Use work-local .agents/work/<category>/<work-slug>/research.md when:
- The research mainly supports one implementation task or work item.
- The findings explain task-specific constraints, alternatives, or plan decisions.
- A future agent should find the research beside
plan.md, prd.md, or progress.md.
Use work-local .agents/work/<category>/<work-slug>/research/<topic-slug>.md when one work item needs multiple focused research notes.
Create or update .agents/research/<topic-slug>.md when:
- Findings are likely to be reused across unrelated work.
- The topic needs multiple sources or non-obvious trade-offs.
- The user explicitly asks to document reusable research.
If work-local findings later become broadly reusable, add a concise promoted synthesis to .agents/research/ and link between the files.
Use this template:
# Research: [Topic Name]
**Date:** YYYY-MM-DD
**Status:** draft | complete
**Question:** [What was researched]
## Recommendation
[Best current answer in 2-4 sentences]
## Key Findings
- Finding 1
- Finding 2
- Finding 3
## Evidence
### [Source or Subtopic]
[Relevant facts only]
## Sources
- [Source Title](url) — What this source established
## Open Questions
- [ ] Include only unresolved questions that materially affect the answer
Status definitions:
draft: Useful findings exist, but material questions remain.
complete: Evidence is sufficient to answer the stated question with confidence.
5. Report Back
Provide the user with:
- A concise summary.
- A clear recommendation.
- Link to the saved research file, if saved.
- Explicit unknowns or open questions.
Definition Of Done
Research is complete when the question has a clear answer or recommendation, supporting evidence is cited, reusable findings are saved when warranted, and the user receives the summary plus any saved file path.