Install¶
Quickstart: let Claude Code do it¶
Paste this prompt into a fresh Claude Code session. It clones the skill, installs the Python dependencies, and walks through credential setup interactively.
Install the
ai-pixel-art-image-generationskill from https://github.com/ianlintner/ai-pixel-art-image-generation into~/.claude/skills/ai-pixel-art-image-generation, install its Python dependencies, then ask whether I want direct OpenAI or Azure AI Foundry forgpt-image-2, and ask for the Gemini API key if I want animations. Don't assume — check which auth path I'm using (OPENAI_API_KEY,az login,DefaultAzureCredential, or a staticAZURE_OPENAI_API_KEY), tell me which shell rc file to export the env vars in, and verify the install by generating a small sprite with--qa.
The manual steps below cover the same ground if you prefer to run them yourself.
Python dependencies¶
rembg downloads a u2net ONNX model on first run (~170 MB); cache it with REMBG_MODEL_DIR if you need to pin the location.
Environment variables¶
# Direct OpenAI:
export OPENAI_API_KEY="<your-openai-api-key>"
# Or Azure AI Foundry:
export AZURE_OPENAI_ENDPOINT="https://<your-foundry-resource>.cognitiveservices.azure.com/"
# Only needed for generate_animation.py (frames 2..N use Gemini).
export GEMINI_API_KEY="<your-gemini-api-key>"
Provider selection is auto: Azure is used when AZURE_OPENAI_ENDPOINT is set, otherwise direct OpenAI is used. Override with --provider openai or --provider azure.
Azure authentication¶
The scripts resolve credentials in this order:
- Azure CLI — run
az loginonce; token cache is reused. DefaultAzureCredential— picks up managed identity, environment vars, or VS Code sign-in.AZURE_OPENAI_API_KEY— static key, last resort.
No endpoint or subscription IDs are baked into the scripts.
Install as a Claude Code skill¶
Clone into your ~/.claude/skills/ directory so Claude Code auto-discovers it via SKILL.md:
git clone https://github.com/ianlintner/ai-pixel-art-image-generation.git ~/.claude/skills/ai-pixel-art-image-generation
Verify the install¶
Generate a small sprite and inspect the QA report:
python3 ~/.claude/skills/ai-pixel-art-image-generation/scripts/generate_sprite.py \
--prompt "orange tabby cat, front view, idle" \
--size 64 --palette auto --transparent-bg --outline palette-darkest --qa \
--output /tmp/cat.png
Expected output ends with a QA table where every hard gate is PASS.