OpenAI Imagegen

OpenAI Imagegen is an official skill for Codex and compatible agents that need to create new raster images or make targeted changes to existing images. The Smithery source originally pointed to the curated imagegen path in the official openai/skills repository. That path is no longer reachable. Checking the current official repository tree identifies the verifiable successor at skills/.system/imagegen. This catalog entry therefore uses the current system path together with OpenAI’s official image-generation documentation. According to the provider, the preferred route is the built-in image tool. A bundled CLI exists only as an explicitly selected fallback and requires locally configured authentication. Those credentials must never be placed in projects, prompts, or catalog copy.

Purpose and audience

The skill is for people who need a bitmap for a website, game, product presentation, prototype, illustration, mockup, cover, infographic, or transparent cutout. It helps the agent first distinguish between creating a new image and editing an existing one. When reference images are supplied, it also establishes whether they provide only style, composition, or mood guidance, or whether one of them is the actual edit target. That distinction protects existing material from unintended changes and makes the requested operation reviewable.

Imagegen is not the right choice when an existing SVG or vector system must be extended exactly, when a logo needs to match an established icon library, or when deterministic output in HTML, CSS, or canvas is sufficient. The skill is not a general image host, an image archive, or a promise of perfect text rendering. For code-native design and exact vector paths, the relevant native technique will usually be more dependable.

Modes and tool selection

The primary source distinguishes a preferred built-in tool mode from an explicit CLI fallback mode. In normal use, the agent should use the built-in image_gen tool and should not switch to a script automatically. If the built-in path is unavailable, the agent should explain that the CLI fallback exists and continue only after the user explicitly chooses it. For CLI use, the source describes generate, edit, and generate-batch operations. The bundled script must not be modified without approval. The skill also explains that built-in output is normally saved under the Codex workspace and that a project-bound asset should be copied or moved into the project afterward.

Before generation, the agent gathers the intended use, asset type, scene, main subject, medium, framing, lighting, palette, materials, exact in-image text, and avoid constraints. A detailed user request should be normalized rather than creatively expanded. For a generic request, the agent may add limited composition or intended-use guidance when it materially helps, but it should not invent unrequested characters, brands, slogans, or objects.

Prompting, variants, and editing

The official guidance recommends a consistent order of scene or backdrop, subject, key details, constraints, and output intent. Short labeled fields are useful for complex requests because they prevent the result from depending on an unstructured mood description. Text inside an image should be supplied verbatim with typography and placement guidance. For edits, the unchanged invariants must be stated explicitly, such as changing only the background while preserving the product and its edges. Each iteration should make one targeted change. That makes it possible to understand whether a difference comes from the new instruction rather than from an unrelated rewrite.

The skill covers photorealistic scenes, product mockups, UI mockups, infographics, illustrations, stylized concepts, and historically grounded scenes. Its editing guidance includes precise object changes, background extraction, lighting or weather changes, style transfer, compositing, and sketch-to-render work. When several variants are needed, the requested differences should be separated clearly. The agent should inspect each result for the subject, style, composition, text accuracy, and preservation of required invariants.

Security, E-E-A-T, and boundaries

This description is grounded in the official OpenAI source at skills/.system/imagegen and the official OpenAI image-generation documentation. According to the provider, the distinction between the built-in tool and an explicitly requested CLI matters; it is not a guarantee of availability, quality, or rights clearance. Images may contain people, brands, confidential references, or copyrighted elements. Before publication, a person should verify that the inputs and outputs may be used, that any required disclosure is present, and that the image does not create misleading or sensitive impressions. A generated image is not evidence that a depicted scene is true.

Local file paths, output directories, and reference images can disclose sensitive information. Unneeded drafts should be retained or deleted under controlled rules. Existing project assets should not be overwritten unless replacement was explicitly requested. Tokens, keys, and other secrets must never be written into examples or output files. A compatible agent may add its own logging, model route, or network services depending on the environment, so the responsible team must review the actual data flow, permissions, and retention separately. Imagegen accelerates visual work, but it does not replace editorial, legal, trademark, or subject-matter approval.