game-art-orchestrator

name	game-art-orchestrator
description	Automates the process of generating game assets matching a specific art style using RAG, Few-Shot visual prompting (via generate_image), and a Human-In-The-Loop VLM evaluation pipeline.

Game Artist Auto Orchestrator

This skill provides a 2-phase pipeline for strictly adhering to a stylistic Generation_DNA and Global rules when generating new game assets.

Workflow Execution Steps

Phase 1: Retrieve & Prompt Engineering (RAG & Few-Shot)

Whenever the user asks you to "draw", "generate", or "create" an asset, BEFORE generating the image, you MUST:

1. Identify the requested style from the prompt. If not explicitly specified, look at the workspace StyleLibrary/ and pick the closest matching subfolder based on semantic meaning.
2. Read the Generation_DNA.md inside that style folder.
3. Automatically retrieve RAG context by executing the retrieval script via terminal: python .agents/skills/game-art-orchestrator/scripts/retrieve_orchestrator_context.py <relative_path_to_style> "<user_prompt>" (For example: python .agents/skills/game-art-orchestrator/scripts/retrieve_orchestrator_context.py StyleLibrary/SciFi "laser blaster")
4. Read the === ORCHESTRATOR CONTEXT PAYLOAD === printed to the console by the script. CRITICAL: Do NOT attempt to manually parse the index JSON files or manually select reference images. You MUST strictly use the text rules and absolute image paths provided by the script.
5. Assemble your final payload by combining the local Generation_DNA.md with the retrieved Global RAG rules and Image paths.
6. Background & Layout Rule Parsing: If the requested asset belongs to an isolated category (e.g., Characters, Items, Obstacles, UI, VFX), you MUST explicitly inject the following constraints into your image generation prompt:
   • "SOLID PURE MAGENTA BACKGROUND (#FF00FF), no gradients, no shadows on the background. Do not generate a floor."
   • CRITICAL: "Generate EXACTLY ONE single asset in the center of the image. DO NOT generate multiple variants, character sheets, split views, or multiple angles. ONLY ONE FIGURE."
7. Execute the generate_image tool (Gemini Nanobanana integration), passing the combined DNA rules into Prompt, and the paths of the Few-shot images into ImagePaths. IMPORTANT: At this stage, append instructions to focus ONLY on Step 1: Sketching & Silhouettes. Do not generate full colors or polish yet.

Phase 2: Sequential Rendering Pipeline (Sketch -> Polish)

Instead of generating multiple variants simultaneously, the orchestration follows a structured 2-step approval process:

1. Step 1: Sketching & Silhouettes Present the generated sketch/silhouette to the user for confirmation. CRITICAL: You MUST embed the generated image directly in your chat response using absolute markdown syntax: ![Image Name](absolute/path/to/artifact.png) so the user can see it. On Windows, you MUST replace all backslashes (\) with forward slashes (/) or use file:/// format so markdown parses correctly. HUMAN-IN-THE-LOOP CHECKPOINT: Stop your execution and ask the user:

   • "Do you approve this sketch/silhouette?"
   • "Are there any structural adjustments needed before we proceed to coloring?" If the user requests adjustments, update the prompt and generate a new sketch. Loop until approved.

2. Step 2: Flat Colors, Shading & Material Polish Once (and ONLY after) the user confirms the sketch, update the generation prompt to apply Flat Colors, Shading, Material Polish and Post-processing while strictly maintaining the approved silhouette. Execute the generate_image tool to create the final render.

3. (Optional) Run python scripts/downscale_image.py <image_paths> if the final image is large and you anticipate Token Overflow when reading it.

4. Adopt the persona in prompts/evaluator-vlm.md. Evaluate the final rendered image against the user's initial prompt, the retrieved Global Rules, and the Evaluation_Rules.json. Provide a Binary Validation Checklist followed by an Aesthetic Score (0-100) and Correction_Guidance.

5. FINAL CHECKPOINT: Present the final image and Evaluator scoring to the user. CRITICAL: You MUST embed the final image directly in your chat response using absolute markdown syntax: ![Image Name](absolute/path/to/artifact.png). Remember to use forward slashes (/) for paths. Ask them:

   • "Do you approve this final render?" (If yes, trigger Phase 3: Asset Finalization & Export).
   • "Should we proceed to Round 2 (Prompt Translation and repolishing) using the correction guidance?"
   • "Would you like to manually override/add any text to the correction guidance before Round 2 runs?"

6. If the user requests Round 2, use the Orchestrator to perform a Prompt Translation on the updated Correction_Guidance. Convert the qualitative feedback into explicit technical parameters (update positive text, inject negative keywords, adjust weights) and re-run Step 2 rendering.

Phase 3: Asset Finalization & Export

When the user explicitly approves a generated variant:

1. Identify the target Project Name from the user's initial prompt or context (e.g., FlappyTrippy). If unknown, verify with the user before exporting.
2. Identify the semantic category of the asset (e.g., Characters, Environments, Items, UI, Obstacles).
3. Construct the absolute export directory path using this professional structure: <workspace>/Assets/Projects/<Project_Name>/GameAssets/<Style_Name>/<Category>/
4. Execute a copy command (do NOT use move/delete) to transfer the approved image artifact from the .gemini/artifacts/ folder into the target directory, ensuring the original remains in artifacts for rendering.
5. Rename the file to a clean, standardized format: [object_name]_[YYYY-MM-DD].png (e.g., laser_pistol_2026-04-20.png).
6. UI Specific Automation: If the category is UI, you MUST automatically perform three tasks:
   • Completeness Check: Review the project's UI Plan or Wireframes documents to verify that NO button, panel, or icon is missing. Cross-reference the generated assets against the full required list.
   • Auto-Resize & Crop: Write and execute a Python script (using PIL) to auto-crop the transparent bounding box (getbbox()) and proportionally resize the UI elements to standard mobile game dimensions (e.g., Icons max 256px, Buttons max 512px, Panels max 1024px) to optimize memory.
   • Integration Guide Generation: Automatically create a <Project_Name>_UI_Integration_Guide.md file in the project root. This document MUST detail how to assemble the generated UI components in Unity (e.g., specifying which elements require 9-Slicing, how to nest Icons inside blank Button bases, and identifying reusable components) to guarantee optimal VRAM usage and responsive UI scaling for developers.
7. Read <workspace>/Assets/Projects/<Project_Name>/GameAssets/Generated_Asset_Catalog.md. Use your code editing tool to append a new row to the table in this file, logging the Category, Asset Name, Target Style, Absolute File Path, and a short 1-sentence prompt description. If the catalog file doesn't exist yet, create it with a standard markdown table header.
8. Confirm completion by providing the user with the final absolute path to their new production-ready asset. CRITICAL: You MUST embed the final exported image directly in your chat response using absolute markdown syntax: ![Image Name](absolute/path/to/exported_asset.png). Make sure to replace all backslashes (\) with forward slashes (/).
9. Integration & Completion: After successful export, you MUST run the /finish workflow to extract any successful prompting techniques or rendering gotchas. Also, remind the user to run /commit to save the newly generated assets.