Write an image brief you can improve one change at a time
You ask for a picture of a team collaborating and get twelve people applauding a laptop. A clearer brief gives the picture a job, a setting and a place to leave empty, and something specific to change when the first result misses.
What you'll make
a six-part image brief, a candidate image, and one focused revision checked in the space where the image will appear.
What you'll need
ChatGPT Images, Adobe Firefly or another image generator in your work account, the size of the space to fill (and where the words go), and permission for any reference you upload. You don’t need to upload people or brand assets to learn the method. About 30 minutes.
Example result.
Job: decorative header for an internal supplier open-day page
Style: flat editorial illustration, teal, sand and grey with lime accents
Framing: 16:9, three people on the right, left third plain for the headline
Exclusions: no lettering, logos, handshakes or dramatic sunbeams
Revision 1: remove the bright burst from the left third. Accepted.1. Give the picture a job and a shape
Write one sentence on where the image will appear and what it should help the reader understand. “A welcoming illustration for an internal page announcing a supplier open day” is a better start than “something professional”.
Decide whether the picture is decorative or meant to show something real. A made-up warehouse can’t show visitors what your site looks like. For that, use a real photograph. Then note the shape: aspect ratio is width against height, so 16:9 is a wide rectangle and 1:1 a square. Check the real template, and reserve the space where the headline will sit before you generate anything. Keep words, logos and dates in the page, not in the picture.
Why this works: most unusable images fail on the job or the space, not on beauty.
What you should see: a one-line job, a shape and the area to keep clear.
2. Write the brief in six parts
Each part is a decision the model would otherwise make for you. Copy this, fill in each line after the colon and delete lines that don’t matter:
Create an image for:
SUBJECT (people, objects or activity):
SETTING (surroundings and relevant background):
STYLE (medium and two or three visual qualities):
LIGHT AND COLOUR:
FRAMING (view, aspect ratio, where the subject goes, which area stays quiet for text I'll add separately):
EXCLUSIONS: No embedded lettering or logos.
If any instruction conflicts with another, tell me which, rather than quietly choosing one.Describe qualities rather than praise: “flat shapes, a limited palette, lightly textured paper” beats “premium”. Use exclusions for the model’s habits you don’t want: glass-walled offices, sunset lighting, people pointing at screens with evident delight. Keep the list short. A long list of things to avoid is harder to follow.
In Firefly or another tool with a single prompt field, turn the six parts into a plain description and set the aspect ratio and style controls in the interface.
Why this works: the framing line ties the picture to your layout, and the conflict line is the cheapest guard against a wasted attempt.
What you should see: a candidate that roughly matches the subject, style and framing, plus a note if the model found a conflict.
3. Name the one thing that missed most
Read the result against the six parts, starting with its job. Can it work in the space? A lovely composition with a face behind the headline is still the wrong composition. Name the biggest problem concretely: “the bright shape sits where the headline goes”, not “it’s not quite right”.
Then look at it at display size and full size. Check an AI-generated image before you publish it covers hands, lettering and edges. A flaw only matters at the size and distance the image will be seen, so slightly blurred background racking on a small header can stay.
Why this works: one named problem gives you one change to make and one thing to judge afterwards.
What you should see: a single sentence describing the main miss.
4. Make one change and recheck the whole picture
In ChatGPT, ask for the edit in the same conversation (or select the area first with the selection tool). In a tool without editing, change the relevant part of the saved brief and run the whole brief again.
Edit this image to fix one issue: [specific problem].
Change [the area or detail] to [the result you want].
Keep [the subject, style, colours and composition that already work].
Do not add text or logos.The “keep” line records your priorities. It doesn’t lock the pixels: OpenAI notes that edits can extend beyond the area you selected. Compare the whole new image with the old one.
Sometimes a crop is quicker than a repair. If a hand is wrong, a tighter crop that removes the hand usually beats three attempts at a better hand. Set a budget before you start: two or three targeted revisions is reasonable. If you’re on the sixth and still negotiating, use stock, a photographer or an illustrator. Knowing when to stop is part of the skill.
Why this works: one change at a time tells you what worked. Five changes at once tells you nothing.
What you should see: the named problem fixed and everything you said to keep still there.
Check the result
- The headline fits in the reserved space with readable contrast, at desktop and mobile crops.
- There’s no lettering, logo or invented signage anywhere in the image.
- Hands, faces and edges hold up at the size readers will see.
- Nothing on the page presents the illustration as your real site or staff.
The usual failure: the image gets the subject and style right and fills the space you reserved, with a decorative burst, a bright window or a stray object exactly where the words go. Spot it by placing the real headline over the image before judging anything else.
Edit this image to remove everything from the [reserved area]. Make that area a plain [background colour] with no objects or pattern. Keep the people, setting, framing and style exactly as they are. Do not add text or logos.[reserved area] is the space you kept for text, such as “left third”. [background colour] is a colour from your palette.
Worked example: a supplier open-day header
Fictional example. An operations manager needs a 16:9 header for an internal open-day page, with three visitors on the right and a headline on the left. It mustn’t pretend to show the real site.
Her first attempt was one line: “a professional image of our supplier open day”. It came back as a glossy wide shot of twelve people in business suits mid-handshake, under a banner carrying the invented word “LOGISTCS”. Impressive, unusable, and wrong in four separate ways.
The six-part brief asked for three fictional adults in hi-vis vests beside one pallet rack, a flat illustration in teal, sand and grey, people on the right two-thirds, the left third pale and quiet, and no lettering, logos, handshakes or sunbeams.
- The failure. The candidate got the people, palette and style right, and put a bright decorative burst behind the headline area.
- The repair. The repair prompt with “left third” and “warm sand” cleared the area and kept the scene. With the headline placed over it, the contrast held on desktop and mobile.
- What she left alone. The clipboard hand had an extra finger, so she cropped slightly tighter rather than asking for a new hand. The blurred racking stayed, because at header size nobody will see it.
Task kit for this job: Download (ZIP) · Reusable instructions
Go deeper
Optional training: keep a reusable house brief
Save the parts that should carry across a series, and leave the subject and setting open:
Purpose and audience:
Style and level of detail:
Palette:
Default shape and text-safe area:
Standing exclusions:
References you're allowed to use, and their source:A shared brief helps with family resemblance but won’t guarantee a matched set. Compare candidates side by side. Keep exact brand colours in the design file. A hex code in a prompt doesn’t produce exact colour values.
Optional training: square crop for the social post
The same header is needed as a 1:1 social image. Don’t crop the 16:9 version. The people would be cut in half. Change only the framing line (“1:1, people centred, top quarter clear for a short title”) and generate again from the saved brief. Then run the same checks. The reserved space has moved, so the usual failure can move with it.
Optional training: save the approved version with its brief
Keep the approved image apart from rejected candidates, with the brief, the tool, any reference permissions and the intended use. That stops the next person mistaking a decorative warehouse for your real premises. Check your organisation’s rules and the tool’s terms before commercial use, and write alt text in the publishing system. If the picture is purely decorative, follow its decorative-image setting instead.