game-art-compiler

name	game-art-compiler
description	Use this skill specifically when the user requests to "compile", "analyze", or "update" an art style based on a directory of asset images. The skill acts as a Data Ingestion pipeline that vectorizes style images using `sentence-transformers` and outputs specialized rules (`Generation_DNA.md` and `Evaluation_Rules.json`) explicitly designed to connect with the `game-art-orchestrator`.

Compile Artist Style V2

Overview

This skill acts as a robust Data Ingestion Pipeline for your game artist workflows. It transforms a loose directory of reference images into a mathematically structured Vector schema (style_index.json) and specific rulesets meant to decouple Generation mechanisms from strict Evaluation heuristics.

Execution Steps

1. Locate Style Directory & Legacy Cleanup

• Identify the exact style name requested.
• Target directory: <workspace>/Assets/GameArtist/StyleLibrary/<style_name>/
• Verify the directory exists via list_dir.
• LEGACY CLEANUP: If a file named DNA_Profile.md exists in this folder, you MUST delete it. It is a deprecated V1 format. Do NOT leave it behind, as it will pollute the system.

2. VLM Semantic Tagging (RAG-Optimized & Dynamic Context)

• Process all valid image files (.png, .jpg, .jpeg) inside the folder.
• Execute a VLM check (e.g. via Gemini/Flash) for EACH image.
• Pass 1: Asset Categorization: The VLM must classify the image into one of: [Character], [Environment], [Prop/Interactable], or [UI].
• Pass 2: Contextual Extraction & Visual DNA: Ask the VLM to produce a highly dense, searchable string (Embedding-Ready). It must explicitly extract:
  1. Visual DNA: Shape Language, Material, Complexity Level / Rarity, and Color Palette.
  2. Game Design Principles (Based on Category):
     • [Character]: Silhouette Readability, Body Alignment, Proportion Rhythms.
     • [Environment]: Guiding Lines, Depth Contrast (Foreground vs. Background), Scale perception.
     • [Prop/Interactable]: Affordance (e.g., sharp/red = danger, round/bright = collectible).
• Also generate a RAG Hook Tag (e.g. "tag: UI_Icon, material: metallic, rarity: epic, visibility: high_contrast").
• Keep a JSON array in memory or disk containing: [{"filename":"...", "semantic_metadata":"..."}]

3. Generate Local Embeddings & Archetype Schema

• Save the JSON array from Step 2 into a temporary file or pass it to our embedding script.
• Execute the script: python scripts/create_embeddings.py <style_directory> (assuming you have passed the captions JSON path or adapted the script to handle inputs).
• The Script will automatically: Use sentence-transformers locally to calculate text embeddings, find the mathematical centroid, assign is_archetype: true to the 2 images closest to the centroid, and output the absolute style_index.json to the target folder.

4. Synthesize DNA and Rule Files

• Collect the macro conclusions from analyzing the full image batch and separate them into two strict files optimized for RAG retrieval and Semantic Search:

  A. Generation_DNA.md Write this file using write_to_file into the style directory. It must use strict nested Markdown headings (for text-splitters) containing the following FOUR pillars:

  • # I. VISUAL DNA: Consistency rules, Shape Language, Color Script (Rarity), and Material Polish.
  • # II. PRODUCT LOGIC: Readability (3-Second Rule), Progression Logic (Level upgrades), The Juice (Animation states), and UI Interaction rules (if applicable).
  • # III. TECHNICAL EXCELLENCE: Mesh / Topology rules, Texel Density, Modular design / Recolor limits.
  • # IV. GAMEPLAY AFFORDANCE & DESIGN LOGIC: Synthesized rules from the dynamic context extraction (e.g., silhouette rules, guiding lines, affordance logic). Include positive anchors and strict negative safeguards under these headings.

  B. Evaluation_Rules.json Write this file using write_to_file into the style directory. It must conform to this schema:

  {
    "semantic_metadata": {
      "supported_types": ["asset", "ui"],
      "vector_keywords": ["list", "of", "dense", "visual", "keywords"]
    },
    "evaluation_criteria": {
      "visual_dna": ["Are shapes consistent?", "Does it match standard rarity palettes?"],
      "product_logic": ["Does it pass the 3-second readability rule?"],
      "technical_excellence": ["Are details readable at thumbnail size?"],
      "gameplay_affordance": ["Does the shape communicate its gameplay function?", "Is it distinct from the background (Visual Hierarchy)?"]
    },
    "forbidden_elements": ["List of anti-patterns"]
  }
  

5. Acknowledge Target & Finish

• Present the updated section explicitly to the user to confirm that the changes were securely registered.
• Remind the user that these global rules will now forcefully apply to the next game-art-orchestrator execution.
• Integration: Automatically run the /finish workflow to document any data ingestion gotchas or conflicts identified during the compilation process.