atmosaero/art-director-loop

style-loop

Discover, refine, and validate a coherent 2D game visual style across characters, items, gameplay elements, backgrounds, and UI.

Vedi sorgente

Come usare

  1. Copia il comando e controlla il pacchetto prima di installarlo nel tuo ambiente agente.
  2. Controlla strumenti, credenziali e ambienti richiesti nel documento originale.
  3. Descrivi input e risultato desiderato; confronta l’output con le fonti.
npx skills add https://github.com/atmosaero/art-director-loop --skill style-loop

Prima di iniziare

  • L’installazione non include abbonamenti, quote API o accesso ai file.
  • Stelle e installazioni indicano attività, non un test indipendente di qualità o sicurezza.
Documento Skill originale

Contenuto dal repository con titoli, esempi, codice, tabelle, link e immagini preservati.

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.

Read at the point of use

WorkReference
Build equivalent test sheets and native-size evidencebenchmark.md
Write each candidate's explicit construction rulesstyle-dna-template.md
Prepare blind packets; score, compare, and critiquejudge.md
Freeze rules and package a standalone handoffstyle-bible-template.md

Keep working artifacts in the game's .style-loop/, separate from this installed skill. Archive versions rather than overwriting judged evidence. The workflow document is the authority for thresholds and counters; the judge document is the authority for scoring.

Present the selected direction, comparable sheets, actionable verdicts, actual-size gameplay evidence, and linked Style Kit. Report either validated, provisional, or unvalidated, with the exact unmet gate when applicable. A beautiful character, a high average, or the user's preference alone cannot certify the visual system.