prototype

Phase 1: Define the Question

Resolve review mode: --review arg > production/review-mode.txt > default lean.

Read concept description. Identify core question the prototype must answer. If vague, state question explicitly before proceeding.

Phase 2: Load Project Context

Read CLAUDE.md for stack/engine/frameworks — prototype uses compatible tooling.

Phase 3: Plan the Prototype

Define in 3-5 bullets: core question, minimum code needed, what to skip (error handling, polish, architecture). Present plan, confirm if scope unclear.

Phase 4: Implement

Ask to create prototypes/[concept-name]/. Every file starts with // PROTOTYPE - NOT FOR PRODUCTION header (question + date).

Relaxed standards: hardcode values, placeholder assets, skip error handling, simplest approach, copy don't import. Run prototype, collect measurable data (frame times, interaction counts, feel).

Phase 5: Generate Prototype Report

Report sections: Hypothesis, Approach, Result (observations not opinions), Metrics (frame time, feel assessment with specifics, action counts, iteration count), Recommendation (PROCEED / PIVOT / KILL with evidence), If Proceeding (architecture, perf targets, scope, effort), If Pivoting (alternative direction), If Killing (why + what instead), Lessons Learned.

Ask to write to prototypes/[concept-name]/REPORT.md.

Phase 6: Creative Director Review

Review mode: solo/lean → skip CD-PLAYTEST. full → spawn creative-director via Task with gate CD-PLAYTEST. Pass report, design question, pillars/core fantasy. CD verdict overrides prototyper recommendation — update REPORT.md if different.

Phase 7: Summary and Next Steps

Output: core question, result, prototyper recommendation, CD final decision, link to report.

• PROCEED: /design-system for production GDD or /architecture-decision
• PIVOT: /prototype [revised-concept]
• KILL: report is the deliverable

Constraints

• Prototype code NEVER imports from production; production NEVER imports from prototypes
• If PROCEED, production implementation from scratch — no prototype refactoring
• Timebox: 1-3 days. If scope grows, stop and simplify the question