Improving One Hundred AI Poster Prompts

Summary

The author describes version 2 of a set of 100 prompts for generating posters with AI image tools. Version 1 often produced cluttered compositions, included every word from an event description as literal imagery, focused too heavily on distinctive details such as ferrets, and invented logos or visual meanings for names such as “The Flood.” To improve the prompts, the author used Claude to develop stronger rules, moved the work into Claude Code, and automated prompt generation and evaluation with OpenAI API credits. Because image generation was more expensive, the first stage was evaluated separately: a language model expanded each prompt, and the resulting instructions were judged against stated goals before images were made. Tests covered five demanding poster styles, including styles directed by the event, styles with their own traditional subject, abstract work, typography-led designs, and a deliberately dense style. The author found that gpt-5.6-sol behaved better than the cheaper gpt-5.4-mini, allowing many model-specific adjustments to be removed. The full process cost about $28, including text evaluation, multi-run image tests, and 200 final images. Version 2 generally produced clearer text hierarchy, more striking posters, and fewer inappropriate subjects, but it still needed retries and continued to favor ferrets. The longer prompts also provide more guidance while giving users less direct control.