Key takeaways
- A prompt only ships after real generations in at least two model families.
- Previews show representative outputs with the model and date noted — never stock stand-ins.
- Each page must clear a written quality bar: breakdown depth, customization, variations.
- Model updates trigger re-tests; stale prompts are updated or removed, never silently left.
The testing loop
Every prompt starts as an experiment, not a caption. A team member writes it, generates with it in at least two model families (for example Flux and Midjourney, or ChatGPT and Stable Diffusion), and compares the outputs against the intended look using the same squint-then-zoom method we teach in How to Recreate an Image With AI. If the prompt can’t survive a different seed or a second model, it goes back for rewording.
Only after the look holds do we write the breakdown. That order matters: the phrase-by-phrase explanations are derived from what we observed the prompt actually controlling, not from what we hoped it would do.
The publish bar
A prompt page ships only when it clears every line of this list:
- At least one real generated example, with the model and date recorded.
- Five or more phrase-level breakdown explanations, written for that specific prompt — never swapped nouns from a template.
- A unique “why it works” section explaining the design logic.
- Three or more concrete customization options.
- Two or more genuine variations (different light, framing, or atmosphere).
- Honest model notes that only claim results for models we actually tested.
Prompts that don’t clear the bar stay on their category page as list entries, or don’t publish at all.
Why the bar exists
Most prompt collections are walls of copy-paste text. The breakdown is our entire value — so the breakdown has to be earned by testing, not invented at the keyboard.
Previews and honesty
Example images on our pages are generated by the team from the published prompt, and each carries its model and generation date in the caption. We don’t use stock photography to stand in for AI output, and we don’t cherry-pick across dozens of rerolls — previews are representative results, and the captions say so. Your output will vary by model, seed and settings; that’s stated on every page because it’s true.
When models change
Model versions shift how vocabulary is interpreted. After any major update to a model we cite, we re-run a sample of affected prompts and refresh the model notes and “last tested” dates. If a prompt no longer performs, we either rework it or pull the page — we’d rather have fewer current prompts than many stale ones. The tool pages carry the same discipline: pricing and capability claims are re-checked each review cycle.
What we won’t publish
- Prompts that name living artists as style targets.
- Prompts designed to recreate a specific creator’s identifiable work or a real person’s likeness.
- Benchmark scores, market-share figures, or ratings we can’t source.
- “Guaranteed identical result” language — variation is a property of these systems, and we say so.
Questions about the standard, or a prompt you think missed it? Tell us — corrections are part of the methodology too.