Prompting

GPTClaudeGemini··▲ 438 copies·updated 2026-07-14

prompting-43.prompt

# Prompting gpt-image-2

GPT Image 2 (April 2026) is not the same animal as the previous generation of image models. The old "magic word" approach — stacking *4K, 8K, ultra detailed, masterpiece, trending on artstation* — is at best useless and often actively makes outputs worse. The new model rewards precise, instructional prompts written like a product spec, not like an incantation.

This document teaches Claude how to write prompts that actually work for this model.

## What changed in GPT Image 2 (vs previous generations)

- **Strong instruction following.** Long prompts (50–300 words) don't lose detail anymore. Be specific.
- **Near-perfect text rendering**, including Chinese / Japanese / Korean. Quote the exact string.
- **Custom resolutions**: any WxH where max side < 3840px and aspect ratio ≤ 3:1.
- **Edit endpoint does precise local edits**: tell it what to change AND what to preserve.
- **Quality tiers exist**, but on this host all tiers cost the same — default `--quality high`.

## The canonical prompt structure

OpenAI's official Cookbook recommends this order. Following it makes the model translate your prompt much more reliably:

when to use it

Community prompt sourced from the open-source GitHub repo jiangmuran/claude-image (MIT). A "Prompting" style prompt — adapt the placeholders and specifics to your task. Imported as-is and not independently retested here, so check the output before relying on it.