uproot-studio

name	uproot-studio
description	Uproot character production studio. Use when creating, tuning, reviewing, QC-ing, or integrating a character/enemy for this project — concept art, base image generation, rig.json runtime values, workbench review, state-capture self-review, enemy kit integration, or status/pipeline questions.

Uproot Studio

Read docs/studio.md first — it is the authoritative workflow shared by every agent (Claude Code, Codex, or human). This skill is a thin router.

Core rules

• Characters here are: emotion via runtime eyes, fixed base image, motion via squash/stretch deformation (no sprite sheets), attachments as separate layers.
• assets/characters/<id>/rig.json runtime section is the single source of truth for runtime values. Never re-introduce literals in src/characters/*Rig.ts.
• brief.md frontmatter holds status and decisions. Decisions land in files, never only in chat.
• Humans never write rig.json/brief.md directly. Workbench tuning and comments land as proposals in assets/characters/review-inbox.md (git-ignored scratch); the AI evaluates each item (accept / adjust / reject with reason), applies accepted changes, and clears the inbox.
• Reuse before generating: check assets/index.json (rebuild with python3 tools/index_assets.py), eye templates in src/characters/eyeEmotionTemplates.ts, burst styles in src/render/feedback/particleBurst.ts.
• Review evidence must come from the real runtime: workbench pages and capture screenshots, not hand-drawn previews.

Reference docs (read before the matching step)

• docs/studio/prompt-rules.md — how to write image-generation prompt packets.
• docs/studio/qc-failures.md — known art failure modes; read before accepting generated art.
• docs/studio/quality-rubric.md — the bar a character must meet before human review.
• docs/studio/eyes.md — the eye system (eye/socket/expression layers, normalized container geometry, template references); read before touching any eye code or gaze values.

Commands

• QC (must be error-free before review): npm run studio:qc
• Self-review screenshots (your eyes): npm run studio:capture -- <id>
• Rebuild reuse index: python3 tools/index_assets.py
• Fit eye containers from base.png: python3 tools/fit_eyes.py <id> (new characters get eye geometry from this, not by hand)
• Generate images: tools/gen_image.sh "<prompt>" <out.png> [style refs...] (codex-subscription image tool; attach existing base.png files as style anchors; obey prompt-rules.md eye-white contract)
• Human review page: npm run dev then /workbench.html (or ?focus=<id>)
• Integration check: npm run build, then screenshot /?encounter=<kind>

Workflow per new character

1. Lock concept: candidates in source/, decision in brief.md (lockedReference, status: reference-locked).
2. Generate base art per prompt packet; accepted art becomes base.png + comparison.png; keep rejects in candidates/.
3. Fill rig.json runtime starting from the closest existing character's runtime section and eye template.
4. Run QC + capture; inspect your own screenshots against docs/studio/quality-rubric.md; iterate until clean.
5. Hand off for human review in the workbench; the human tunes values in-page (preview only) and submits proposals + comments to review-inbox.md; the AI processes the inbox and selectively applies changes to rig.json.
6. Integrate: write src/game/enemies/<id>Kit.ts modeled on an existing kit, register it in src/game/enemies/index.ts, map a room theme in CombatRoomScene, build, capture /?encounter=<kind> evidence into playtest/, advance status.