Research Skill
Investigate a technical problem efficiently and persist what you learn so the team never re-investigates the same topic twice.
1. Scope the topic
Extract from the user's message (ask only if genuinely unclear):
- Topic slug: kebab-case folder name, e.g.
bluetooth-audio-routing - Stack: platform/language/framework
- Symptom: observed vs expected behavior
- Known: hypotheses already tried
2. Prepare the workspace
Run scripts/init-research.sh {topic-slug} to create the docs folder.
3. Launch 3 default agents in parallel
Each agent gets a one-line angle plus the shared context. Read reference/agent-prompts.md
when you need the exact prompt wording.
- Codebase analysis (Explore agent) — find relevant code, gaps, file:line references.
- Root-cause theory (web agent) — why it happens at the OS/framework/protocol level.
- Solutions & validation (web agent) — native APIs, libraries, community-confirmed fixes.
Spawn the optional 4th/5th agent only if the user explicitly asks for library comparison or cross-platform API deep-dive.
4. Synthesize
Create exactly these two files:
docs/{topic}/NOTAS.md— raw findings, snippets, URLs, quick observations.docs/{topic}/RESEARCH_CONSOLIDADO.md— structured master doc.
Use reference/consolidated-template.md only when you need the full template.
Optional detail files (03_android_api.md, 04_librerias.md, etc.) are created only when
findings justify them.
5. Finalize
Run scripts/finalize-research.sh {topic-slug} to validate the master doc exists and append
an iteration row. Then report to the user:
- Root cause (with file:line if applicable)
- Top 3 non-obvious findings
- Files created
- Suggested next step
Keep the chat report concise; the detail lives in the docs.