How to use the Codex transparent asset generator agent skill

What the skill makes the agent do

The skill does not draw images itself. It tells the agent to write a detailed image prompt, call the built-in Codex image generation tool to create the picture, and then confirm the result. The confirmation step runs a small Python helper that checks whether the PNG actually contains transparent pixels rather than only having an alpha channel.

The agent also keeps the original generated file, records the prompt and file hashes, and can copy finished assets into a project's folders. If the built-in image tool is not available, the skill tells the agent to stop and explain the blocker instead of switching to another image provider.

• Writes a complete asset prompt covering material, silhouette, padding and transparent openings.
• Generates one distinct asset per call when separate files are needed.
• Runs a read-only check that reports format, size, alpha range, transparent percentage and visible bounds.
• Keeps original files and saves the prompt and SHA-256 hashes for each asset.
• Can copy assets into a project and compare hashes to confirm the copies match.

Install the skill

You can add the skill in one of two ways. Download the skill ZIP from this page and unzip it into your personal skills folder at ~/.claude/skills/. Alternatively, put it in a single project by unzipping it into that project's .claude/skills/ folder.

If you use the Claude apps, you can instead upload the ZIP under Settings → Capabilities → Skills.

After unzipping, the skill folder should contain agents/openai.yaml, scripts/inspect_png.py and SKILL.md.

• Personal install: ~/.claude/skills/
• Project install: the project's .claude/skills/ folder
• Claude apps: Settings → Capabilities → Skills, then upload the ZIP

What you need before you start

The skill depends on two things. The first is the built-in Codex image generation tool (named image_gen.imagegen in the skill) being available in your agent environment. The skill states that it cannot run without this tool. The second is Python 3 with the Pillow library, which the inspection script needs.

The skill does not let you choose a model. The built-in tool exposes no model selector, so the skill cannot guarantee which model produced an image. The name "GPT 2.5 Flare" is just the workflow's name, not proof of a model.

• Built-in Codex image generation tool available in the agent environment
• Python 3 with the Pillow library for scripts/inspect_png.py
• No model choice: the agent will not claim which model ran

How to trigger the skill with natural requests

Ask for transparent artwork directly and name the kind of asset you want. The skill's description mentions requests for transparent PNG ornaments, game sprites, emblems and UI graphics. You can also refer to the skill by name, such as using the /codex-gpt-image-2-5-flare command or asking for the GPT 2.5 Flare workflow.

Be specific about the asset's purpose, size or aspect ratio, and whether it needs to stack over a background. The skill asks the agent to request dimensions in the prompt when the tool has no dedicated size parameter, and then to check what comes back.

• "Make a transparent PNG crest for my inventory screen, centered, with transparent holes through the design."
• "Create one corner ornament for a fantasy UI panel, an asymmetrical L at top left, mostly empty toward the center."
• "Generate a transparent health-bar sprite for my game and verify the alpha channel."
• "Use /codex-gpt-image-2-5-flare to make a set of isolated emblem graphics for my project's assets folder."

Example workflow from request to checked file

When you send a request, the agent first reads your request and any reference images, then writes the prompt. For an isolated asset it uses a standard instruction asking for one object as a PNG with genuine alpha transparency, with no painted checkerboard, solid backdrop, floor or rectangular shadow, and no text or logo unless you ask for it.

After generation, the agent looks at the image, then runs the inspection helper on the returned file. The helper's command has this form, with the paths replaced by your own:

• Two compositions are described in the skill: a corner (an L shape anchored at one corner) and a crest (a centered emblem using about 85% of its canvas).
• The helper rejects files that are not PNG, and with the transparency option it also rejects images that are fully opaque or entirely invisible.
• The helper only reads the file. It never edits or converts pixels.

Safety notes: scripts and network

The skill runs a script. The bundled helper scripts/inspect_png.py is described as read-only. The skill states that it does not use the network. Because it runs code on your machine, review the script before installing it if you work in a sensitive environment.

The skill also limits what the agent may do with files: it keeps originals, does not bypass your image pipeline, and does not infer deployment permission from image generation. Repository commit and deployment rules still apply.

• Runs scripts: yes (scripts/inspect_png.py, read-only)
• Uses network: no, according to the skill's description
• Originals are kept; project copies are checked by hash
• Repository commit instructions are followed; generation alone does not authorise deployment

Tips for better results

Ask for one asset at a time. The skill generates one distinct asset per call when separate files are needed, and it advises against sprite sheets unless you request one. Mirrored corners can share one texture, so ask for a single corner and reuse it.

Check transparency against the real background. The helper's numbers do not prove that enclosed gaps are see-through. Place the asset over the background it will use, or view a checkerboard preview. If the result is wrong, ask the agent for an edit using the original as a reference rather than chroma-keying the background away.

• Request one corner or one crest per call, and say which size it will display at.
• Keep interactive buttons and labels out of the artwork; the skill keeps them in the UI.
• Ask for vector or hand-drawn SVG only if you want something else: the skill avoids substituting vector assets.