Every card carries a four-clause prompt you can paste, a diagram of the framing it asks for, and the part most prompt lists leave out what to expect back, and how dependably. Fourteen of these land first time. Six are here specifically because they drift, with the workaround attached.
We have not generated these images and we are not going to publish invented pictures as though a model produced them. Each thumbnail is a small drawing of the framing and arrangement the prompt asks for useful for picking the right prompt at a glance, and honest about what it is.
The result information is in words instead, and it is more useful anyway: what to expect tells you what the prompt tends to produce, and the reliability chip tells you how much of it you should verify. Those ratings come from where image generators generally drift rendered text, exact counts, specific real subjects, extremities and consistency across a set not from a scored test.
Every prompt also carries a change first line, because the useful skill is not finding a better prompt but knowing which clause to adjust when the one you have misses. The beginner's guide covers the loop that goes around these.
Backgrounds, mood and concepts ask for atmosphere rather than accuracy, which is exactly what approximation is good at. These are the jobs worth giving it.
Nothing specific to get wrong. Glance at it and use it.
A real object, a count, or an extremity in frame. Sixty seconds of looking.
Included deliberately, each with the workaround that actually gets you there.
A smooth abstract backdrop in dusty rose and slate, matte risograph grain, empty through the centre with detail only at the edges, flat even light.
Expect: a usable canvas on the first try. The “empty centre” clause is what makes it a backdrop rather than a picture.
Torn cream paper laid flat with one rough edge, scanned analogue texture, filling the frame edge to edge, soft overhead light.
Expect: a tileable-feeling surface. Good under text you add yourself later.
A café interior thrown far out of focus, warm 35mm film photography, wide horizontal frame with the centre uncluttered, late afternoon window light.
Expect: warmth and depth with nothing legible in it which is the point, since nothing legible means nothing to get wrong.
Coarse natural linen weave, macro photography, filling the frame at a slight diagonal, raking side light to raise the texture.
Expect: the light clause does nearly all the work here without it you get flat fabric, with it you get texture.
A weathered slate surface with fine cracks, documentary photography, seen from directly above, square crop, a single soft light from the left.
Expect: a moody flat-lay base. “Directly above” is worth stating or you get a three-quarter angle.
Overlapping translucent shapes in teal and amber, cut-paper collage, centred with generous margins on all sides, flat studio light.
Expect: clean geometric abstraction. The margins clause keeps it from bleeding to the edges, which matters if it is going behind something.
Rain running down a window pane with a blurred street beyond, cinematic photography, cool desaturated palette, close on the glass, the city soft behind it, scattered streetlight through the water.
Expect: atmosphere in one shot. Everything beyond the glass is meant to be illegible, so nothing can be wrong.
Low fog lying across an empty field with a treeline behind, quiet landscape photography, muted greens, wide, horizon low in the frame, pale sun just clearing the mist.
Expect: stillness. “Horizon low” leaves sky space, which is useful if something goes over the top.
An empty open-plan office with one lamp still on, restrained editorial photography, wide, shot down the length of the room, a single warm pool of light in cool ambient dark.
Expect: a strong single-light image. The contrast between one warm source and cool surroundings is what carries it.
A worn kitchen counter with steam rising from a pot, warm documentary photography, film grain, mid-distance, counter running across the frame, cold blue daylight from a window on the right.
Expect: the warm/cold contrast in the light clause is doing the emotional work. Drop it and the image goes flat.
A dark storm sky with one bright break over the sea, dramatic landscape photography, wide, sky taking two thirds of the frame, a shaft of sunlight hitting the water.
Expect: high drama, dependably. Stating the sky/sea ratio is what stops it centring the horizon.
Stacked blocks rising in a steepening curve, clean flat vector illustration, three-colour palette, left to right across a wide frame, plenty of space above, flat even light, no shadows.
Expect: a clean slide graphic. Ask for no shadows explicitly or flat vector tends to acquire them.
Many streams converging into one narrow channel and few emerging, minimal isometric illustration, flowing left to right, the pinch point centred, flat even light.
Expect: the idea reads immediately. Do not ask for labels add those yourself.
Scattered dots on the left resolving into a single clean line on the right, minimal editorial illustration, two colours, wide, the transition happening at the centre, flat light on a plain ground.
Expect: a reliable metaphor image. Two colours keeps it from becoming decorative.
Each of these puts something in frame that approximation gets nearly right: a real object, a count, a hand. They are still good prompts they just need the four-zone check before anything goes out.
A soft abstract field in deep green and cream, no text anywhere, grainy print texture, square, with a clear empty band across the middle, flat even light.
Expect: a good backing plate and occasionally some invented squiggle that looks like lettering. Saying “no text anywhere” helps and does not guarantee.
A single bold symbolic object on a saturated ground, no lettering, graphic poster illustration, limited palette, square, object centred with clear margins, hard directional light for a strong shadow.
Expect: strong thumbnails, because one object at square crop survives being shrunk. Check for stray glyphs.
A simple seasonal motif with nothing written on it, flat illustration, three colours, wide banner, subject pushed to the right, left third left empty, flat light, no shadow.
Expect: the asymmetry clause is the whole trick it reserves the space your title goes in.
A decorative border motif in ink and gold, centre left plain, letterpress illustration, vertical, pattern at the top and bottom only, even light with a slight paper sheen.
Expect: a frame you can set details into. Ornate borders are where invented lettering likes to hide.
A chipped enamel mug on a worn wooden sill, 35mm film photography, warm muted palette, close crop, centred, shallow depth of field, soft late-afternoon side light.
Expect: a convincing generic mug. It will not be your mug for that, photograph yours and edit from it.
An open notebook with a pen laid across it, pages unmarked, clean product photography on a plain ground, directly overhead, notebook filling most of the frame, soft diffused light, minimal shadow.
Expect: good composition, and handwriting-like marks on the pages unless you say unmarked and sometimes even then.
A plain unlabelled glass bottle on a stone plinth, editorial product photography, cool neutral palette, centred, low camera angle, plinth filling the lower third, one hard light from the side for a long shadow.
Expect: a strong mockup base. Say unlabelled, because invented labels are invented text.
A person seen from behind standing at a tall window, quiet cinematic photography, mid-distance, figure small against the frame, strong backlight turning them to silhouette.
Expect: the safest way to put a person in an image backlit and turned away means no face and no hands to go wrong.
Two hands resting on a mechanical keyboard, close documentary photography, desaturated, tight crop from just above, keyboard filling the frame, soft light from a screen off to one side.
Expect: good atmosphere and an even chance of the wrong number of fingers. Count them, every time.
An original character of your own describing: a tired lighthouse keeper with a knitted hat, warm storybook illustration, waist-up portrait, centred, soft lamplight from below.
Expect: a charming one-off, and a different person every time you ask. Getting the same character twice is group seven's problem.
Ask for a named product and you get a convincing relative of it close enough to look right, wrong in the details, and misleading if anyone takes it for the real thing. If a specific object matters, photograph it and edit from the photograph, which is also faster.
Real people are a firmer line. Generating a recognisable likeness of someone raises problems no prompt solves, so the people prompts here are either turned away from the camera or invented outright. For a real person, use a real photograph of them.
These are here because people try them anyway. Each one names the workaround that actually gets the result, which is usually not a better prompt.
A corner shopfront with a large blank sign board above the window, street photography, overcast palette, straight on, sign in the upper third, flat overcast light.
Expect: asked for lettering, you get confident nonsense. Asked for a blank board, you get a board you can set real type onto.
Exactly five smooth river stones in a row, minimal still-life photography, straight on, evenly spaced across the frame, soft top light.
Expect: four, or six. Counts are approximated rather than followed, and small numbers are no safer than large ones.
The same lighthouse keeper from before, now on a harbour wall, warm storybook illustration, full figure, mid-distance, low evening light.
Expect: a relative of the first character. Words cannot carry a likeness between generations that is what references are for.
A simple line chart shape with no labels or numbers, flat vector illustration, two colours, wide, axes in the lower left, flat even light.
Expect: asked for data, you get invented figures rendered as text wrong twice over. Asked for a shape, you get artwork.
A generic red-brick civic building with a clock gable, architectural photography, neutral palette, straight on, building filling the frame, flat overcast light.
Expect: name a real landmark and you get something that resembles it and is wrong in every detail a local would notice. Describe a generic one and it is simply a building.
A small repeating leaf motif across the whole frame, block-print illustration, two inks, flat and even across the frame, no focal point, flat light.
Expect: a lovely pattern whose edges do not meet. Tiling needs exact edge continuity, which approximation does not provide.
The pattern across all six: when a request needs something to be exactly so, the tool is the wrong instrument for that part of the job. It is still the right instrument for the rest of it what approximation means in practice.
No, and they are not presented as such. Each card carries a composition diagram we drew by hand to show the framing the prompt asks for. Publishing invented pictures as genuine output would be the easiest thing to do here and the least honest.
It is our assessment of how dependably a request of that shape lands, based on where image generators generally struggle: rendered text, exact counts, specific real subjects, extremities and consistency across a set. It is a guide to what to check, not a measured success rate.
Because four is enough and it stays adjustable. Subject, style, composition and light cover what an image prompt can usefully specify, and keeping them separate means you can change one clause at a time instead of rewriting a paragraph the reasoning in full.
Yes, that is what they are for. They are also meant to be edited swap the subject clause and keep the style and light clauses, which are the parts that carry a look between images.
Because naming one produces a convincing near-miss rather than the thing itself, which is misleading at best, and because generating likenesses of real people raises problems no prompt solves. Supply a real photograph instead when a specific subject matters the longer answer.
Reliably, you do not. Rendered text is a known weak point across image generators, so the working method is to generate the picture with a deliberately blank area and add the words yourself afterwards in any editor.
Change one clause and try again rather than rewriting the whole thing, since one change per attempt is the only way to learn what your wording does. If the result was nearly right, edit that image instead of generating a fresh one.
No. Generation is not reproducible and there is no seed to record, so the version you save is the only copy of that result.
Keep the style and light clauses identical and change only the subject, then use your best result as a reference image for the rest of the set. Describing a look repeatedly gets you near it; showing it gets you onto it.
No. Image generation is free with rate limits that most people never reach, so working through a batch of these is not something to ration pricing and limits.
The reusable part of any of these is the style and light wording. Keep those fixed, swap the subject, and you have a set that looks like it belongs together.
Independent guide not affiliated with Meta. These prompts are untested suggestions written to a structure, not recipes with verified results; what you get back will differ every time.