ChatGPT Image Prompting Cheat Sheet

OpenAI publishes no prompt formula — but they do rewrite your prompt before generating it, and that changes how you iterate.

TL;DR — THE SHORT VERSION

OpenAI publishes no prompt formula for images. Their docs cover implementation, not craft.
Your prompt is automatically revised before generation. The rewritten version comes back in a revised_prompt field.
So read revised_prompt before rewriting. The mismatch may be the revision, not your wording.
No negative prompt field. Exclusions go in the conversation, in plain language.
Documented weak spot: precise text placement and layout-sensitive composition. Prompting does not fix it.

VENDOR-DOCUMENTED

This sheet quotes the vendor's own published prompting guidance.

The formula

No vendor formula published — conversational, iterative

Template:

[what the image is] [subject in detail] [setting and light] [style] [what to avoid, stated plainly]. Then refine in follow-up messages, one change at a time.

Worked example:

A documentary photograph of a ceramicist centring clay, north-facing studio window at dusk, muted palette, shallow depth of field. Avoid glossy over-lit skin and stock-photo posing.

Breakdown — what each part does

The request

Say what kind of image it is first — photograph, illustration, diagram. The model behaves differently for each.

Subject and setting

Ordinary descriptive prose. This model handles natural language better than keyword strings.

Exclusions

Stated conversationally in the same message or a follow-up. There is no separate field.

Iteration

Follow-up messages, changing one thing at a time. This is the model's actual strength.

Does it take a negative prompt?

No negative prompt field. Conversational only.NO

Write exclusions as plain sentences — "avoid glossy over-lit skin" — in the prompt or a follow-up. Keep them few; a long list of avoidances competes with the description and this model is already rewriting your text before it generates.

What not to do

Do not assume the model saw what you typed — check revised_prompt when using the API.✕

Do not fight the layout limitation with more prompt detail; OpenAI states it struggles with precise element placement.✕

Do not paste keyword-soup prompts written for diffusion models. This one reads prose.✕

Do not change five things at once when refining — you lose track of which one worked.✕

THE REVISION BEHAVIOUR

This is the single most consequential undocumented-in-guides fact about the model. If output does not match your prompt, the cause may be a rewrite you never saw. On the API the rewritten text is returned to you; in the chat interface it is not, which is why the same prompt can behave differently in the two places.

The checklist

Before you send it:

1 — Said what kind of image it is?□

2 — Written in prose rather than keywords?□

3 — Exclusions stated plainly and kept few?□

4 — Checked revised_prompt if on the API?□

5 — Refining one element per turn?□

SOURCE

OpenAI image generation guide, developers.openai.com. Checked 25 Aug 2026.

Model versions and vendor documentation both move. Re-read the source before relying on a specific number. Errors are logged at corrections.

Every other sheet in this set: all model cheat sheets.