A production image brief names the subject, composition, intended placement, required and forbidden elements, aspect ratio, and any exact text that must be checked later. Describe constraints that matter to the asset rather than stacking style adjectives. If editing an existing image, identify which parts must remain stable and which may change. Generated pixels require visual review for text accuracy, identity, accessibility, and unintended content. Do not treat a model's textual assurance that the image is correct as proof.
Make this comfortable
Image prompts: specify the visual contract and inspect the pixels
Decision in practice
A logistics dashboard needs a hero image for a tablet-width layout. The brief asks for a warehouse sorter on the left, a clear quiet area on the right for interface text, neutral lighting, and no visible company marks. The team generates three candidates, then checks the actual crop at 16:9 and a narrow mobile crop. One candidate places a pallet in the supposed quiet area; another invents a badge with unreadable lettering. Both fail despite satisfying most of the written brief. The accepted image is stored with its prompt version and review notes.
Asset: warehouse dashboard hero.
Composition: sorter left, right third visually quiet.
Format: 16:9 source; verify mobile crop.
Exclude: logos, badges, embedded copy.
Review: subject placement, crop, legibility, unintended marks.Performance and operating cost
Generating multiple candidates increases model cost and review time roughly with candidate count. Large output dimensions add processing and storage cost; resizing a small image does not recover missing detail. A fixed acceptance checklist makes the review faster than repeated subjective revisions. Keep exact copy in the interface when possible rather than trusting image-rendered letters. Store the accepted asset and its dimensions so later page changes can be checked against the crop and loading budget.
Common Mistakes
- Do not accept an image from its prompt alone.
- Do not depend on generated lettering for critical copy.
- Do not ignore the final responsive crop.
Connected lessons
- Prompt engineering applications
- Prompt Engineering
- Multimodal prompts: separate what an image shows from what it suggests
- Evaluation sets: measure the failure cases that matter
- SQL prompts: generate a query under database-enforced limits
- Audio prompts: preserve speakers, timestamps, and uncertain words
- Project: produce a checked multimodal incident brief
- Applied prompt engineering decisions
Continue with: Image prompts: verify claims against named regions and crops.
Continue with: Video generation prompts: specify one shot's motion and evidence.
Continue with: Screenshot prompts: separate pixels from behavior.
prompt engineering
tutorial