What our own prompt library actually looks like
Not advice — measurement. Every figure below comes from counting this site's own published prompts and generations, dated to when it was measured. Re-run periodically as the library grows; each page carries the date its numbers are as of. Questions is the place for general how-to answers instead.
What does a good AI image prompt actually look like?
We read all 782 published prompts in this library and measured what they have in common: a median of 74 words, five parts in a fixed order (subject, setting, lighting, camera, style), and a negative prompt that is almost entirely about suppressing text. Below are the actual prompts and the actual images they produced, taken apart line by line.
Based on 782 published prompts and 1,134 generated images in this library.
read →What is the best way to write image prompts?
Write in five parts, in this order: subject, setting, lighting, camera, style. Every one of the 782 prompts published here is stored that way, and the reason is mechanical — the front of the prompt gets the most attention, so decisions that define the picture go first and finish words go last.
Based on 782 published prompts, component-level structure and copy-rate data.
read →How do you prompt image generators — and does the same prompt work everywhere?
The prompt travels; the interpretation does not. We ran 68 identical prompts through gpt-image-2, FLUX 2 Pro and Nano Banana 2: the three engines chose a different composition for 23 of them and a different dominant colour for 35. Here is what that looks like in practice, and what to change per engine.
Based on 68 prompts run through all three engines (204 generations).
read →Looking for a how-to instead?
These pages measure what already works here. For general technique — settings, fixing artifacts, keeping a character consistent — see the questions hub.