Media and prompt rules
Quality rules for FluxonLab prompts, generated media, provider limits, and review gates.
FluxonLab uses prompts as production inputs, not disposable text. A media or content prompt should be specific to the project, sized for the provider, and reviewable after the run.
General prompt rules
- State the subject, audience, use case, and output format.
- Include only the context the model needs.
- Avoid secrets, private credentials, and unverified claims.
- Make negative rules explicit: no fake logos, no unreadable generated text, no invented UI labels unless the job requires them.
- Keep provider limits in mind. If a provider has a character limit, compress the prompt without losing intent.
- Archive the final prompt next to the generated output so future review can explain where it came from.
Image prompts
Image prompts should describe the scene, composition, style, aspect ratio, and intended placement. Website images and social images are different jobs:
| Target | Prompt focus |
|---|---|
| Website | Product evidence, hero/section fit, negative space, clean crop behavior. |
| Social | Clear post angle, strong visual hook, caption/overlay space, no generated pseudo-text. |
Generated images should not rely on model-written text for logos, labels, or brand marks. When text must be exact, create it deterministically outside the image model.
Video prompts
Video prompts should keep motion concrete and short. For short generated clips:
- describe the first frame, camera motion, subject motion, and ending frame;
- keep text minimal and specify where it appears;
- prefer short readable phrases over dense captions;
- avoid asking the model to create complex UI copy frame by frame;
- archive the prompt, provider, duration, model, and output path.
If a video contains broken characters, decide whether deterministic overlay/cropping can fix it or whether the clip needs regeneration.
Review gates
Media review should check relevance, visual quality, text legibility, brand safety, duplicate outputs, provider errors, and whether the result matches the target channel. Rejected outputs should move to a reject archive with a manifest entry and, when useful, a regeneration reason.