style-loop
Style Loop
Treat style as a production system: one explicit visual grammar must survive multiple asset categories, actual gameplay sizes, and previously unseen content. Deliver a reusable Style Kit with evidence of its limits.
Start and route
Read workflow.md for execution, persistent state, budgets, and exit gates. Choose from the user's intent:
- Discover: no established direction, or an explicit redesign. Compare 3–5 structurally different candidates (default 4), shortlist, refine, provisionally lock, stress-test, then finalize.
- Refine: preserve an existing identity. Capture current assets/screens and baseline DNA; compare that baseline with two targeted alternatives unless the user already selected a direction. Then use the same refinement and validation gates.
- Validate/extend: an existing Bible or locked style is the starting point. Audit its rules and evidence, run the common benchmark and new-content test, and propose changes only where failures require them. Do not reopen divergence merely because another style seems prettier.
- Resume: recover the phase, artifact versions, unresolved issues, and remaining budgets from the project ledger. Do not restart counters or discard failed rounds.
Capabilities
Need project file access, visual inspection, and a way to produce visible test artifacts: image generation, existing assets, or suitable 2D rendering tools. Use available image generation for raster exploration; follow its tool/skill instructions. Code-native vector/UI work may use the project's own renderer. A text prompt is not a rendered artifact.
Use a fresh independent judge subagent when delegation and image access are available. The judge procedure and fallback are in judge.md. Tool unavailability does not justify invented inspection, scores, images, or independence. Without visual evidence, deliver an explicitly unvalidated plan/DNA and stop at the documented capability exit.