<objective>
Comprehensive research on HOW to implement a phase before planning.
This is for niche/complex domains where Claude's training data is sparse or outdated. Research discovers:
- What libraries exist for this problem
- What architecture patterns experts use
- What the standard stack looks like
- What problems people commonly hit
- What NOT to hand-roll (use existing solutions)
Output: RESEARCH.md with ecosystem knowledge that informs quality planning.
</objective>
<execution_context>
@/.claude/get-shit-done/workflows/research-phase.md
@/.claude/get-shit-done/templates/research.md
@~/.claude/get-shit-done/references/research-pitfalls.md
</execution_context>
<context>
Phase number: $ARGUMENTS (required)
Load project state:
@.planning/STATE.md
Load roadmap:
@.planning/ROADMAP.md
Load phase context if exists:
Check for .planning/phases/XX-name/{phase}-CONTEXT.md - bonus context from discuss-phase.
</context>
<process>
1. Validate phase number argument (error if missing or invalid)
2. Check if phase exists in roadmap - extract phase description
3. Check if RESEARCH.md already exists (offer to update or use existing)
4. Load CONTEXT.md if it exists (bonus context for research direction)
5. Follow research-phase.md workflow:
- Analyze phase to identify knowledge gaps
- Determine research domains (architecture, ecosystem, patterns, pitfalls)
- Execute comprehensive research via Context7, official docs, WebSearch
- Cross-verify all findings
- Create RESEARCH.md with actionable ecosystem knowledge
6. Offer next steps (plan the phase)
</process>
<when_to_use>
Use research-phase for:
- 3D graphics (Three.js, WebGL, procedural generation)
- Game development (physics, collision, AI, procedural content)
- Audio/music (Web Audio API, DSP, synthesis)
- Shaders (GLSL, Metal, ISF)
- ML/AI integration (model serving, inference, pipelines)
- Real-time systems (WebSockets, WebRTC, sync)
- Specialized frameworks with active ecosystems
- Any domain where "how do experts do this" matters
Skip research-phase for:
- Standard web dev (auth, CRUD, REST APIs)
- Well-known patterns (forms, validation, testing)
- Simple integrations (Stripe, SendGrid with clear docs)
- Commodity features Claude handles well
</when_to_use>
<success_criteria>