PROMPT CRAFT

The braces were never the point.

Everyone is writing image prompts as neat objects now, and the pictures do look better. The reason is not the one being shared, and knowing the real one makes the tidiness optional.

THE SHORT VERSION

Structure is a discipline for the person writing. It is not a language the picture understands.

•  OpenAI's own guide for its image models says minimal prompts, descriptive paragraphs, structured objects, instruction-style prompts and tag lists can all work well, and to use whichever is easiest to maintain.

•  Google teaches the same craft in the opposite costume. Its image guidance is written almost entirely in full descriptive sentences.

•  So why do the neat ones look better? Because a form with empty fields makes you fill them. The discipline is doing the work, not the punctuation.

•  Underneath every image prompt that works, in any shape, sit the same four groups. Five finished examples are written out below.

Braces have been having a moment. Open any feed of image work and half the prompts arrive as tidy objects, keys quoted, commas lined up, and the pictures attached to them are genuinely good. The implication travels without ever being stated: write it like this and the model takes you more seriously.

It is a reasonable thing to believe, and the people sharing it are not being dishonest. They are just crediting the other half of what they did.

This is a question you can settle by reading rather than arguing, because both companies publish how they think their own models should be spoken to.

OpenAI's guide for its image models addresses the format question head on. It says minimal prompts, descriptive paragraphs, structured objects, instruction-style prompts and tag lists can all work well, and to use whichever is easiest to maintain. For anything you run repeatedly it recommends a skimmable template over clever syntax. That is a house telling you the container is your convenience, not the model's requirement.

Google's image guidance looks like the opposite and means the same thing. It is written almost entirely in flowing description, teaching you to ask for a photorealistic shot of a subject in a setting and then fill that in with real words. No keys, no braces, and the results people get from it are not worse.

Two houses with opposite habits of presentation, one shared assumption underneath. The model needs the information. It does not care what shape you hand it over in.

If the shape does not matter, the honest question is why structured prompts so obviously produce better pictures. They do. That part of the trend is real and worth keeping.

The answer is that a form with named, empty fields will not let you skip anything. Sit down to write a prompt as a flowing sentence and you will write the parts you find interesting, which is usually the subject, and stop when it sounds finished. Sit down to fill in a field called lighting and you cannot leave it blank without noticing you left it blank. The structure is not talking to the model at all. It is talking to you, and what it says is: you have not told it where this happens.

Which is why the comparison people run proves less than they think. A careful structured prompt against a lazy sentence is not a test of format. It is a test of care, and care wins, as it was always going to.

Take the braces off a thorough prompt and it stays thorough. Put them around a vague one and you get a vague picture with tidy punctuation.

Strip the punctuation off any image prompt that works and the same four groups are underneath it. Keeping them in a settled order is the one piece of formatting advice both houses give out loud, because then a change to one part reads as a change to one part.

The scene comes first: where this happens, what time of day, what sits behind everything. Then the subject: who or what, and what they are doing, which is the group people write in most detail and still under-specify. Then the details that decide the look: the light, the lens or the distance, the materials, the colour, the mood.

Then the constraints, and this is the one that goes missing. The shape of the frame. Any words that have to render correctly. Anything that must stay out. What the picture is actually for. Almost nobody writes this group down, because the model you have used for months has been quietly guessing it from habit and getting it right.

That is also why a prompt that has served you for a year can fall apart the first time you paste it into something else. Nothing in the text got worse. It was always incomplete, and the old model was covering for you. Write those four as keys in an object, as labelled lines, or as two well-built sentences. What matters is that all four are there, and that you will still understand them in six months.

Five different jobs, five different genres, no braces anywhere. Each one names the scene, then the subject, then the look, then the constraints, and each ends by saying what the picture is for, which is the line that quietly settles a dozen decisions the model would otherwise make on its own. Paste one and change the nouns.

ONE  ·  EDITORIAL PORTRAIT

A corner office with north-facing windows in late afternoon, city rooftops soft and out of focus behind the glass. A woman in her fifties in a charcoal wool blazer sits on the edge of the desk, turned toward the window, caught mid-thought rather than posed, one hand resting flat on the desktop. Soft directional daylight comes from camera left with a bounce card opening the shadow side, shot on an 85mm lens at f/2 with the focus held on the eyes, in a muted palette of greys and warm skin with a single note of oxblood. Frame it vertical at four by five, head and shoulders down to mid-torso, with generous empty space above her head for a headline to sit. Render no text in the image, no jewellery and no visible brand marks. This is for the leadership page of an annual report.

TWO  ·  PRODUCT ON SEAMLESS

A putty-grey studio sweep with no horizon line anywhere in frame. A matte amber glass dropper bottle stands upright and dead centre, cap on, label square to the camera. One large softbox sits above and slightly behind it with a narrow strip light to camera right, laying a soft elliptical shadow that pools toward the viewer and a clean specular highlight down the left edge of the glass, shot on a 100mm macro at f/8 so everything from cap to base stays sharp. Frame it square, with the bottle filling the middle sixty percent and even margins on all four sides. Include no props, no water droplets, no hands, and no reflection of the studio in the glass. Colour needs to be accurate enough to match a printed swatch. This is for a marketplace listing on a white background.

THREE  ·  INTERIOR

A small kitchen in a pale Scandinavian style, mid-morning in early spring, the window at the left of frame opened a little onto a grey sky. The room is the subject and it is empty of people, with a linen tea towel folded over the oven rail and a wooden board holding two lemons near the sink. Cool daylight rakes across a pale oak counter and warm light bounces up off a terracotta floor, shot wide on a 24mm lens at f/5.6 with the verticals kept straight, so it reads as lived in rather than styled. Frame it horizontal at three by two, camera at standing eye height, with the whole counter run visible from one edge of the frame to the other. Include no people, no appliances carrying visible brand marks, and no text. This is for the header image of a rental listing.

FOUR  ·  FOOD

A dark walnut table beside a window on an overcast day. A shallow bowl of hand-cut noodles sits in clear broth with steam still rising, chopsticks resting across the rim, and a small dish of chilli oil just behind it and softly out of focus. A single window light comes from behind and slightly left so the steam catches and the surface of the broth glints, with a bounce card at the front keeping the shadows open, shot on a 50mm lens at f/2.8 from a forty-five degree angle. Frame it vertical at four by five with the bowl sitting on the lower third and clean quiet space across the top. Include no hands, no printed menu, and no garnish beyond what is named here. This is for a restaurant social post where a headline will be placed over the top third.

FIVE  ·  CINEMATIC STILL

A rain-soaked provincial bus terminal at night, fluorescent tubes running overhead and puddles across the concrete holding the light. A young man in a thin windbreaker stands still with a duffel bag over one shoulder, looking off toward something outside the frame while other passengers move past him. The colour temperature is mixed, cold from the overhead tubes and warm from a sodium-lit food stall at the right, shot on a 40mm anamorphic lens at f/2 with him held sharp and the moving crowd smeared by a slow shutter, carrying fine film grain and deep blacks that still hold detail. Frame it widescreen at two point three nine to one, placing him on the right third so he looks across the open left side. Include no on-screen text, no legible signage, and no visible logos. This is a reference frame for a storyboard.

Read the last two sentences of each one again. That is the constraints group, and it is doing more work than every adjective above it.

Paste this into a text model, then paste one of your working image prompts underneath it. It is written to find gaps rather than to improve anything, because a pass that also reworks your taste is not a check.

Below is an image prompt I use. Sort every piece of information in it into four groups: the scene and background, the subject and what it is doing, the details that decide the look such as light and distance and materials and colour, and the constraints such as frame shape, any text that must render, anything that must stay out, and what the picture is for. Print the four groups with my own words under each, changing nothing. Then list what is missing from each group, and for every missing item say what a model would most likely assume if I left it out. Do not improve my wording, do not add adjectives, and do not write a new prompt.

The instruction to name the likely assumption is the working part. That list is a map of everything the model has been deciding on your behalf, and it is the shortest route to a prompt that reads the same to anything you point it at.

THE NEXT WORKING NOTE

Putting together a shape kit: one image prompt written five ways, as an object, as labelled lines, as a paragraph, as a tag list and as two sentences, all carrying the same four groups, with the same picture beside each and a note on what each shape costs you when you come back to it in six months.

Want it when it ships? Reply with send me the shape kit and I will get it to you.

A QUESTION FOR YOU

Do you write your image prompts as structure, or as sentences?

Reply and tell me which, and whether you arrived at it deliberately or just never changed. I suspect most people inherited their shape from whichever guide they read first, and I would rather hear it than assume it.

If this was useful, forward it to whoever keeps your prompt library.