Save evidence, not magic words
A prompt cannot be separated from the reference assets, model, settings and result that made it useful. A repeatable library stores the full production recipe and the conditions under which it succeeded.
The minimum prompt record
- Name and intended visual job.
- Output type and selected model.
- Aspect ratio, duration, resolution or image quality.
- Prompt and negative or exclusion language.
- References, character and world used.
- Successful output and known failure examples.
- Credits, retries and reviewer notes.
Organise by job to be done
- Product reveal and feature demonstration.
- Character introduction and continuity.
- World establishment and environment transition.
- Vertical hook and copy-safe close.
- Storyboard shot and campaign bridge.
- Image reference, first/last frame or editing operation.
Mark fixed and variable fields
Fixed fields protect the subject, factual offer, world rules and exclusions. Variable fields invite a controlled change such as hook, camera, lighting or ending. This turns the library into a system that can generate measured variants without losing brand truth.
Version prompts after real output
- Draft: written but not yet tested.
- Tested: produced a technically valid result.
- Approved: passed brand, rights and delivery review.
- Model-specific: depends on one route or reference mode.
- Retired: no longer reliable or no longer permitted.
Keep the library small enough to trust
Archive duplicates and prompts that cannot be tied to an approved result. Review high-value entries whenever the model catalogue changes. A library of twenty proven recipes is more useful than hundreds of generated prompt lists with no output evidence.
