Labelling AI content
What has to be disclosed, by whom, and from when.
Read the guide →By task
Image prompts
Image models reward specificity about what is in the frame and how it is seen. They punish adjective piles.
Five decisions, in order
A useful image brief answers these. Leave one out and the model picks for you, usually toward whatever is most common in its training data.
Subject
What is actually in the frame, and what it is doing. Be concrete about number, age, clothing and posture if they matter, because vagueness here is where stereotypes fill the gap.
Composition
Shot type and viewpoint. Close-up, wide, eye level, from above. This is the single biggest lever on whether an image feels intentional.
Light
Direction, quality and time of day. Soft overcast light, hard low sun from the left, a single lamp. Light is what makes an image look photographed rather than generated.
Style and medium
Photograph, illustration, diagram, and in what tradition. Naming a medium is more reliable than naming a living artist, which raises obvious rights questions.
Constraints
Aspect ratio, negative space for text, colours to avoid, and what must not appear.
A wide shot of two colleagues reviewing printed drawings on a workshop bench.
Eye level, slightly off-centre, the bench filling the lower third.
Late afternoon light from a high window on the left, soft shadows, dust visible in the beam.
Documentary photograph, muted colours, shallow depth of field.
16:9. Leave clear space in the upper right for a headline. No text, no logos, no visible faces.Editing needs a preservation list
When you ask for a change to an existing image, the model has no idea which parts you consider settled. Say so explicitly, or it will helpfully improve things you were happy with.
Change only the sky, to overcast with soft even light.
Preserve exactly: the building, its colour and materials, the position of every window, the people and their clothing, the framing and crop.
Match the new lighting to the existing shadows on the ground.
Do not sharpen, recolour or restyle anything else.Text in images is still the weak point. Rendering of words has improved a lot, but it remains the most common place an otherwise good image falls apart. If the text matters, generate the image with space for it and add the text yourself.
The part most image guides skip
If you publish AI-generated or AI-manipulated imagery in the EU, labelling obligations start applying on 2 August 2026. The duty to disclose material that appears authentic falls on whoever uses the system, not only on whoever built it, and it applies even where there is no intention to deceive. Artistic and satirical work gets lighter treatment.
In practice this is a workflow decision rather than a legal puzzle. Decide how you mark AI imagery, apply it consistently across every channel, and make sure the people publishing know the rule.
Multimodal prompting
Briefing work on images, PDFs, audio and video you supply.
Read the guide →Method and sources
How we treat vendor claims about model capability.
Read the guide →