Worked AI Image Prompt Examples, Clause by Clause

These are worked image prompt examples built for diagnosis: each shows the prompt, describes the result, and identifies the clause doing the work. Nearly every other page of ai image prompt examples has the same shape. Here is a prompt, here is a picture, copy the prompt. Meta's guide does ten, one site does 260 to copy and paste, another does sixty-nine, a marketplace advertises four thousand, and there's a tested-fifty list aimed at content creators. Some of them are genuinely good prompts.

What none of them tell you is which part of the prompt caused which part of the picture. So you can copy one, get something nice, and be completely stuck the moment your subject is different.

These eight are written the other way round. Each one is a full prompt I ran, a description of what came back, and the single clause I'd point at as load-bearing. Then at the end there's a deletion test where I took one of them apart a clause at a time to find out which words were actually holding it up. Small samples throughout, three or four runs each, which I'll flag as I go because a single generation tells you nothing.

How To Read These

One thing at a time, and reroll before you believe anything.

If you need subjects to practise on before you need finished prompts, start with the ideas for learning prompt control. That page generates difficult briefs; this one diagnoses what makes a completed image prompt work.

Generation is random, so a version that comes back better once is not evidence. I run each variant at least twice and preferably three times before deciding a word did something, and I still catch myself getting excited about what turned out to be a different seed.

The classes below come from the six-class split I use for image prompts generally, which is set out in ai image prompt. Photograph, illustration, flat vector, 3D render, text-forward, diagram. The vocabulary that steers one is inert in the others, so the examples are grouped that way rather than by subject.

One, A Photograph With A Reason To Exist

A bakery counter fifteen minutes after closing. Two unsold loaves left on a wire rack, one paper bag folded flat beside the till. Low warm light from a single fixture over the counter, everything beyond it in shadow. Medium shot, 50mm, slight downward angle. 35mm film photograph, muted colour, fine grain. 4:5. No text, no signage, no people.

Three runs, two I'd keep. What came back read as a real end of day rather than a product shot, and the thing doing that is the phrase about two unsold loaves, because a specific small quantity implies a whole day that already happened.

The load-bearing clause is "fifteen minutes after closing". It sets the light, the emptiness, and the mood in five words, and I'd bet it does more than the lighting sentence that follows it. Time of day as a scene-setter is enormously efficient.

Two, A Flat Mark With No Light In It

A flat vector symbol of a folded paper boat. Two colours only, deep navy and off-white, no gradients, no shadows, no light source. One closed shape, thick uniform stroke, generous negative space around it. Centred on a plain background. Square.

Four runs, one clean keeper and two nearly-there. The two nearly-there ones had a faint drop shadow that crept in anyway, which is the model's habit and not a failure of the prompt exactly.

The load-bearing clause is "no light source". Saying "flat" isn't enough, since flat is a word attached to thousands of soft branding mockups, whereas explicitly denying that light exists in this image pushes it toward actual vector work. Colour count matters almost as much.

Three, An Editorial Illustration

An editorial illustration of a person asleep at a kitchen table beside a cold cup of tea. Heavy tapered ink line, uneven weight. Two flat tones for shading, no gradient. Limited palette, four colours, one warm accent on the cup. Visible paper texture. Simplified anatomy, hands suggested rather than drawn. 3:2.

Three runs, two keepers, and the failures were both anatomy.

The load-bearing clause is "hands suggested rather than drawn", which is a trick I stole from real illustration rather than from prompting. Hands are where generated illustration falls over, and instructing simplification in advance is far more effective than fixing it afterwards. It also happens to be what a lot of illustrators do anyway.

Four, A Product Render

A studio product render of a matte ceramic desk lamp, warm grey glaze, brushed brass fittings. Soft key light from upper left, large softbox, subtle rim light separating it from the background. Seamless mid-grey backdrop. Camera at product height, slightly off axis. Sharp focus on the switch detail. Ultra clean, no dust. 1:1.

Four runs, three usable, which is my best keep rate in this whole list and I think that's because product renders are close to the centre of what these models saw a lot of.

Load-bearing clause is "camera at product height, slightly off axis". Left unspecified, renders default to a slightly high three-quarter view that reads as a catalogue thumbnail. Google's own product mockup template gives camera angle its own bracket, which suggests they found the same thing.

Five, Text That Has To Be Right

A poster design for a music night with the text "TWO SETS" in a heavy condensed sans, placed in the upper third. Deep red background, single off-white ink. Bold geometric shapes below the type, no photographic elements, high contrast. Print poster feel with slight ink texture. 2:3.

Four runs. One where the lettering was correct, two where it was nearly correct in a way that's worse than being wrong, and one where it invented a third word.

The load-bearing clause is the short string in quotes. Two words, eight characters. OpenAI's own documentation says the model "can still struggle with precise text placement and clarity" even after significant improvement, and my hit rate agrees with them. Keep strings tiny, put them in quotes, say where in the frame they sit, and if the wording genuinely has to be right, set the type yourself afterwards.

Six, A Diagram That Isn't Lying

A simple labelled cross-section of a French press, drawn as a technical illustration. Uniform thin line weight, no shading, no perspective, straight-on elevation. Four labels only, plain sans lettering, leader lines to the plunger, filter, spout, and handle. White background. 4:3.

Three runs and I would not publish any of them without redrawing the labels.

That's the honest answer for this class. Models produce confident, attractive diagrams with lettering that looks right at thumbnail size and falls apart at full size, and the danger is that a diagram carries an authority a photograph doesn't. The load-bearing clause is "four labels only", which at least kept the invention contained. I'd treat generated diagrams as layout drafts and never as finished reference.

Seven, A Three Panel Sequence

Make a 3 panel comic in a loose ink and wash style. A person waits at a bus stop in the rain, checks an empty timetable, then sits down on the wet bench. Same character across all three panels, same coat, same umbrella. Muted palette, three colours. Panel borders thin and hand drawn.

Three runs, one I'd use.

Google publishes a sequential art template that's much shorter than mine, essentially a style plus a scene, and I added the continuity instructions myself because the first attempts drifted. The load-bearing clause is "same coat, same umbrella", since naming specific repeated objects holds a character together better than describing the character does. Faces still drift. Objects hold.

Eight, A Still Built To Become A Clip

A narrow city street at dawn, wet asphalt, nobody in frame. A single shop sign glowing at the far end, everything else in blue pre-sunrise light. Locked-off wide shot, low camera height, 35mm. Muted colour, fine grain. 16:9. No text, no vehicles.

Three runs, two keepers, and this one exists to be a first frame rather than a final image.

Load-bearing clause is "nobody in frame". An empty establishing shot is the easiest thing to animate afterwards, because nothing in it can walk wrong, and the motion instruction can be pure camera. That workflow, generating the still first and animating from it, is the single biggest improvement I made to my own video work, and it's covered in ai video prompts.

The Deletion Test

I took example one apart, removing one clause at a time and running each version twice. This is the table I wish the prompt banks published.

Two things stand out. The time-of-day clause and the quantity clause were carrying more than either lighting or lens, which is the opposite of what I'd have guessed a year ago. And the aspect ratio removal did not crop, it recomposed, which catches people out constantly.

The angle row is the useful negative result. It did almost nothing, twice, so it's a candidate for deletion in every prompt of this shape. Ten minutes of this on your own prompt teaches you more than the next two hundred examples in any collection.

What The Eight Have In Common

Every load-bearing clause in that column is a constraint rather than a description. Not an adjective anywhere. That's the pattern I'd take away, and it's consistent with what I've found across everything else I've tested, where the words that do the most work are the ones that close down possibilities rather than the ones that pile on qualities.

The text and diagram rows are the honest low points. I'd rather show a keep rate of zero than pretend otherwise.

Questions

How many examples do you need to learn from? Fewer than any bank offers. Eight or ten with the reasoning attached beat four thousand without it, which is the whole argument I keep making.

Do these prompts work on every model? The structure does. The exact behaviour won't, since aspect ratio handling, text rendering, and default aesthetics differ per system. Treat them as shapes rather than as recipes.

Why is your keep rate so low? Partly standards, partly that I generate locally on an M4 Pro, so an extra attempt costs me electricity and I'm not incentivised to settle. If you're paying per image you'll accept more, and that's a rational response to a different price.

Where are the text prompt examples? Separate piece, since the levers are completely different. Before and after edits across text, image, and video are in ai prompt examples.

What about composition specifically? Placement, negative space, and palette get their own treatment in ai artwork prompts.

Try This One Thing

Take your best prompt and delete one clause. Run it twice. Then put it back and delete a different one.

Half an hour of that and you'll know which third of your prompt is decorative, which is a thing no example collection can ever tell you, because they don't know your subject. The framework these examples hang off is in chatgpt prompts.