GPT Image 2.5 Prompting
Turn the user's intent and optional reference images into a GPT Image 2.5-ready prompt. Keep API settings separate from the visual prompt. Generate an image only when the user explicitly asks for execution and a suitable image-generation tool is available.
This skill distills the official OpenAI image prompting guide. Treat that guide and the image generation documentation as authoritative when behavior or parameters change.
Route the request
Identify the mode before drafting:
- For text-to-image creation, read references/generate.md.
- For editing, restyling, compositing, translation, cutouts, or any request with input images, read references/edit.md.
- For a follow-up that modifies an earlier output or preserves a recurring character, read references/refine.md.
- When selecting Flare versus Sunburst, returning request settings, or executing generation, also read references/api-parameters.md.
Read only the references needed for the current request. A request may need an editing reference followed by the refinement reference, but it does not need every mode by default.
Preserve the request
- Retain the user's subject, action, style, exact text, composition, aspect ratio, and relationships among reference images.
- Do not replace a chosen model, rewrite supplied copy, or add subjects, logos, text, or stylistic themes that change the intended result.
- Add visual detail only when it clarifies an underspecified request without changing its meaning.
- If an edit depends on an image that is not available, ask the user to attach it. Do not fabricate reference-image details.
Build the prompt
Use the shortest maintainable format that fully expresses the request. Plain sentences, descriptive paragraphs, labeled sections, and JSON-like structures are all acceptable; there is no required magic syntax.
For a complex request, organize the prompt around the applicable parts below:
- Result and use: Name the deliverable and its intended use.
- Subject and action: Describe visible subjects, poses, gaze, scale, and interactions.
- Composition: State framing, viewpoint, hierarchy, placement, and negative space.
- Visual treatment: Describe medium, materials, lighting, palette, texture, and atmosphere.
- Exact text: Quote required copy and specify placement, typography, and occurrence count.
- Changes: For edits, state exactly what should change.
- Preservation constraints: State what must remain unchanged.
- Exclusions: Name unwanted extra text, logos, watermarks, objects, or changes when relevant.
Treat camera and lens terms as appearance cues, not guarantees of exact physical simulation. Do not add a generic negative-prompt block when targeted exclusions are sufficient.
Return the result
Return the visual prompt in a fenced text block. When relevant, follow it with separate sections for:
- Reference roles: Map each supplied image to its purpose and destination.
- Request settings: Provide model, size, quality, background, and output format outside the prompt.
- Checks: Call out exact text, factual labels, identity, product geometry, or transparency that must be inspected in the result.
Do not mix model choice, API parameters, prices, or access requirements into the visual prompt.
Execute only when requested
- If the user asks only for a prompt, stop after returning the prompt and any relevant settings.
- If the user asks to generate or edit an image and an appropriate tool exists, use the selected prompt, reference images, and request settings with that tool.
- After execution, inspect the result against the stated requirements. Do not claim exact text, factual diagrams, preserved identity, or true transparency without checking.
- For revisions, make one focused change at a time and repeat the critical preservation constraints.
Final checks
Before returning or executing a prompt, confirm that:
- Every user-specified requirement is present and no invented requirement changes the intent.
- Required text is quoted exactly and the prompt forbids extra text when appropriate.
- Every reference image has an explicit role.
- An edit separates changes from preserved details.
- API settings are valid and remain outside the visual prompt.
- The response does not imply that writing a prompt has already generated an image.