Prompting: One Brief, Many Dialects

011. The brief comes first

Seven slots

A complete brief answers seven questions. They are the same questions an art director answers when commissioning a person.

Part 1 showed why this matters. An underspecified prompt is completed with whatever was most common in the training data. Ask only for “a woman repairing a bicycle” and you will get a young woman, outdoors, in daylight, centred, smiling at the camera, holding a spanner near a bicycle that needs no repair. Every one of those is a default you did not choose.

You do not have to fill every slot every time. You do have to know which ones you left empty, because those are the ones the model decided.

The vocabulary of pictures

Models learned from captions written by people who make pictures, so the working vocabulary of photography and illustration is unusually effective. It is also more precise than adjectives. “Dramatic” means little. “Lit from one side by a single low lamp” means one thing.

• Framing: close-up, medium shot, wide shot, full-length, overhead, low angle, over the shoulder, profile, three-quarter view.
• Lens and focus: wide-angle, 35mm, 85mm portrait lens, macro, shallow depth of field, everything in focus, motion blur.
• Light: soft window light, overcast, golden hour, backlit, rim light, hard midday sun, a single practical lamp, neon, candlelight.
• Medium: documentary photograph, studio product shot, gouache, linocut, watercolour, flat vector illustration, isometric diagram, pencil sketch, 3D render.
• Finish: film grain, matte, high-key, low-key, muted palette, limited to three colours.

A useful habit for anyone who does this often is to collect the terms that reliably produce a visible effect in the model you use, and ignore the rest.

Describe the style, not the artist

Naming a living artist is the fastest way to a style and the one to avoid, for the reasons part 6 gave. It is also less controllable than it looks: a name brings the artist’s subjects and compositions along with the manner. Describing the manner, as “loose gouache, visible brush strokes, flat shapes, limited earthy palette”, travels between models and leaves no one’s name in your working files.

People, specifically

If a picture contains people, the defaults are where bias lives. Say who is in the picture: age, build, clothing and, where it matters to the use, ethnicity and gender. Then generate a batch and look at all of it. A single image tells you nothing about what the prompt tends to produce, and what it tends to produce is what will end up across a campaign.

022. Why models want different prompts

Two facts from part 1 decide what a model can read.

The first is its text encoder. CLIP treats a prompt as something close to a bag of concepts, with a hard limit of 75 usable tokens. A T5 encoder or a full language model reads sentences and keeps track of which adjective belongs to which noun.

The second is the captions the model was trained on. A model trained on web alt text has seen millions of captions like “sunset beach photo canvas print”. A model trained on machine-written descriptions has seen millions like “A wide photograph of a beach at sunset. In the foreground, wet sand reflects an orange sky”. Each learned to respond to what it saw.

How to tell which you have

For a hosted service, read the prompting guide. For an open model, the model card names its text encoders, and a community fine-tune’s page nearly always shows example prompts. Those examples are the most reliable evidence there is: they show the dialect the fine-tune was trained in.

Failing that, test it. Take one brief, write it as keywords and as two sentences of prose, fix the seed as part 1 described, and generate both. Then add a relationship that keywords cannot express, such as “the lamp is on her right and the window is behind her”. A model that gets the relationship right from prose is a prose model.

The cost of the wrong dialect

• Prose into a keyword model: the 75-token window fills with “a”, “the”, “which is” and “in the”, and everything after it is cut off or weakened. The relationships you wrote so carefully are not understood anyway.
• Keywords into a prose model: it works, but the result is generic, because nothing says how the pieces relate. The model has the ability to follow a detailed description and was given a shopping list.
• Tags into a base model: tag vocabulary it never saw, such as “1girl”, is noise at best.
• JSON into a keyword model: the field names themselves, “subject”, “lighting”, “camera”, become words in the prompt. You may get a picture of a camera.

033. Keywords and tags

The keyword prompt

The running example, for a first-generation model:

documentary photo of a middle-aged woman repairing a bicycle wheel, grey hair tied back, canvas apron, cluttered bicycle workshop, evening, single warm work lamp, deep shadows, blue window light in background, 35mm, shallow depth of field, film grain, amber and teal

And its negative prompt:

smiling, looking at camera, daylight, outdoors, cartoon, illustration, blurry, deformed hands, extra fingers, watermark, text

The rules behind it are few.

• Order is priority. Earlier tokens carry more weight. Lead with the medium and the subject, and put finish and colour last.
• Stay inside the window. CLIP reads 75 tokens at a time. Tools accept longer prompts by cutting them into chunks of 75 and encoding each separately, so a phrase that straddles a boundary is split in two. Keep a prompt to one chunk where you can. Where you cannot, most tools have a keyword that forces a chunk break, so that you choose where the cut falls.
• No grammar, no negation. “Without a helmet” contains the word helmet. Anything unwanted goes in the negative prompt.
• One concept per phrase. “Red scarf, green coat” leaks colour between the two less than a long clause does, though part 2 explained why it still leaks.

Weights

Tools for open models let you turn the emphasis on any phrase up or down. In the most widely used syntax, round brackets raise attention and a number sets the factor:

documentary photo of a woman repairing a bicycle wheel, (single warm work lamp:1.3), (deep shadows:1.2), cluttered workshop, (smiling:0.6)

A bare pair of round brackets multiplies by 1.1, and square brackets reduce. Useful values sit between about 0.5 and 1.5. Beyond that the image degrades in the same harsh, overcooked way as excessive guidance, because the effect is similar: one direction is being exaggerated.

Weights are a feature of the tool, not the model. They work by scaling the encoder’s output for those tokens, and the syntax differs between tools. They do little or nothing on prose models, whose encoders were not built for it.

What goes in the negative prompt

Part 1 explained what a negative prompt is: the thing guidance pushes away from. That makes it a steering tool, not a bin for anxieties. Long boilerplate negatives copied from galleries are mostly superstition, and every token in them pulls the image away from something, often something harmless.

Use it for three things: the defaults you know the model will reach for (smiling, looking at camera), the wrong medium (cartoon, 3D render, when you want a photograph), and known defects of that model. Keep it short, and remember from part 1 that few-step and guidance-distilled models ignore it entirely.

Quality words

Early prompts were full of phrases such as “highly detailed”, “8k” and “trending on” a named art site. They had some effect, for an unglamorous reason: in scraped captions those words sat beside polished, professional images. The effect was always modest and is weaker still in newer models. One or two do no harm. A dozen use up a fifth of the window.

Tag-trained models are the exception, as the next section explains.

Tag prompts

Many illustration and anime fine-tunes were trained on images from image-board sites, where volunteers label every picture with tags from a controlled vocabulary of many thousands of terms. For those models the tag vocabulary is the language, and ordinary description works poorly.

A tag prompt has a customary order: quality tags, the number and kind of subjects, the character, their features and clothing, the pose and action, the framing, the setting, and then style.

masterpiece, best quality, 1girl, solo, mature female, grey hair, ponytail, apron, repairing, bicycle, bicycle wheel, looking down, focused, upper body, from below, workshop, indoors, evening, lamp, warm lighting, dark background

Three points are specific to this dialect.

• Quality tags are real here. These models were trained with tags marking how highly each image was rated, so “masterpiece, best quality” or a model’s own rating tags genuinely select for its better training images. The model’s page says which tags it expects, and using another model’s will not work.
• Use the exact tag. “Ponytail” is a tag. “Hair tied back” is not. The tag for one female subject is “1girl” whatever her age, which is why “mature female” follows it. Tools offer tag auto-completion for this reason.
• The tags are the limits. Anything the vocabulary has no tag for is hard to get. Composition and relationships between two characters are the usual casualties, and part 6’s control images are the usual answer.

a [stone cottage:glass pavilion:0.4] in a pine forest, morning mist

044. Descriptive prose

The prose prompt

The running example, for a prose model:

A documentary photograph of a woman in her fifties truing a bicycle wheel in a small, cluttered workshop in the evening. She has grey hair tied back and wears a canvas apron, and she is looking closely at the rim, not at the camera. A single warm work lamp on her right lights her face and hands and leaves the rest of the room in deep shadow. Behind her, a small window shows dim blue dusk. Medium shot from slightly below bench height, with her left of centre. 35mm lens, shallow depth of field, slight film grain, a palette of amber and teal.

It carries the same seven slots as the keyword version. The difference is that it can say how they relate: the lamp is on her right, the window is behind her, she is left of centre, and she is not looking at the camera.

How to write it

• Most important first. Order still matters. One publisher’s guide says so directly and gives the order to use: main subject, then key action, then the critical style, then essential context, then secondary details. If the image must be a photograph, say so in the first five words.
• One idea per sentence. Plain declarative sentences are read more reliably than one long sentence with many clauses. This is the register the training captions were written in.
• Say where things are. Left, right, behind, in the foreground, in the top third. Spatial language is what prose models gained over keyword models, and most prompts do not use it.
• Name the light source. Where the light comes from fixes more of the look than any style word.
• Be concrete. “A chipped enamel mug” outperforms “a nice mug”. Specific nouns carry more information than adjectives of approval, and “beautiful” carries none.
• Stop when the brief is complete. The same guide suggests 30 to 80 words for most work, with longer prompts for complex scenes. Beyond that, details begin to compete, and late ones are dropped.

What prose models do not do

• Negation. “No people in the background” still risks adding people. Describe what is there: “the background is an empty brick wall”.
• Negative prompts, often. Many current models have none. One major publisher’s guide says flatly that its model does not support them, and to describe what you want. Where a tool still shows the box, part 1 explained how to tell whether it does anything.
• Bracket weights. Emphasis is given the way a writer gives it: by putting a thing first, giving it more words, or saying “the main focus of the image is”.

Lettering

Current prose models letter short text well, and the conventions are consistent across them. Put the exact words in quotation marks. Say what the text is on and where it sits. Describe the lettering as you would to a sign-writer: “bold condensed sans-serif capitals, white on red”.

A hand-painted wooden sign above the workshop door reads "MARLOWE CYCLES" in cream serif capitals on dark green, with "Repairs while you wait" in smaller script beneath it.

Keep it short, check every letter, and expect part 2’s advice to hold: for anything that matters, add real type afterwards in a design tool.

Exact colours

Some current models accept colour codes in the prompt, tied to a named object: “the apron is #2F4F4F”. It is more reliable than colour words and useful for brand colours. Treat it as a strong hint and not a guarantee, and check the result with a colour picker.

055. Structured and JSON prompts

What it is

During 2025 a style spread in which the prompt is a data structure. The running example:

{
  "scene": "A small, cluttered bicycle workshop in the evening",
  "subjects": [
    {
      "description": "A woman in her fifties, grey hair tied back, canvas apron",
      "action": "Truing a bicycle wheel, looking closely at the rim, not at the camera",
      "position": "Left of centre, medium shot"
    }
  ],
  "lighting": "A single warm work lamp on her right. Deep shadow elsewhere. Dim blue dusk through a small window behind her",
  "camera": {
    "angle": "Slightly below bench height",
    "lens": "35mm",
    "depth_of_field": "Shallow"
  },
  "style": "Documentary photograph, slight film grain",
  "color_palette": ["amber", "teal"],
  "mood": "Quiet concentration"
}

This works only where the text encoder is a language model, or where a language model sits in front of the image model. Such a model has read an enormous amount of JSON and understands that the value of “lighting” describes the lighting. There is no standard schema. At least one publisher documents a recommended set of fields for its model, close to the one above, and most models of this kind will accept any sensible field names.

What it does and does not buy

Be clear-eyed about this, because the style is often oversold. One publisher that documents a JSON schema also says that its model understands both formats equally well, and that the choice should depend on your workflow. For most models that is the position: for a single image written by hand, well-organised prose does as well.

There is one real exception, and it shows where things may be heading. An open model released in June 2026 by a design-tool company, built for typography and layout, was trained only on structured JSON captions, with fields for each element’s position, styling and colours. For that model JSON is not a convenience but the native dialect. Even there you need not write it by hand: by default a language model rewrites a plain request into the JSON the model expects, which is the prompt rewriting that section 6 describes. The lesson is the one from section 2. Find out what the model was trained to read.

The benefits are real, and they are engineering benefits.

• Fields keep attributes apart. Two subjects, each with its own description, position and action, are less likely to swap clothes and colours than the same content in one paragraph.
• Templates become trivial. A brand’s lighting, camera, style and palette are fixed fields. Only the subject and scene change per image. Software fills the slots.
• Changes are reviewable. A difference between two versions of a JSON prompt shows exactly which field changed. That is the discipline of treating prompts as code, which part 3 of the practical AI module covers.
• Iteration is cleaner. Section 8’s rule is to change one thing at a time. With fields, one thing is one field.
• It forces a complete brief. An empty “lighting” field is visible in a way that a missing sentence is not. For many people this is the largest benefit, and a checklist would achieve the same.

Where it fails

On any model with a CLIP encoder the braces, quotation marks and field names are all just tokens, and they displace real content from a 75-token window. On tag-trained models it is meaningless. Even on models that read it, deep nesting and dozens of fields dilute the prompt in the same way a 300-word paragraph does.

The middle path

If the model does not read JSON, the discipline is still available. Write prose in a fixed order with one sentence per slot, or use labelled lines:

Subject: a woman in her fifties, grey hair tied back, canvas apron.
Action: truing a bicycle wheel, looking closely at the rim.
Setting: a small, cluttered bicycle workshop, evening.
Light: one warm work lamp on her right, deep shadow elsewhere, blue dusk in a window behind.
Camera: medium shot from below bench height, subject left of centre, 35mm, shallow depth of field.
Style: documentary photograph, slight film grain, amber and teal.

Most prose models handle this well. Keep the structured version in your own systems, and generate whichever dialect the target model needs from it. That arrangement also means a change of model is a change of template and not a rewrite of every prompt.

066. Hosted services and assistants

Someone else writes your prompt

Part 1 mentioned that some assistants rewrite your prompt. The practice began when one vendor found that its image model, trained on long machine-written captions, performed far better on long detailed prompts than on the short ones people actually type. So a language model was placed in front to expand every request.

The consequences are worth knowing.

• Short requests work, and come back embellished. Ask for “a woman repairing a bicycle” and the rewriter adds a setting, a time of day, a style and, often, deliberately varied demographics. The defaults are now the rewriter’s and not the image model’s.
• Your own long prompt may be rewritten too. If you have written a complete brief, say so: “use this prompt exactly as written”. Some services honour that, some only partly. Many will show you the prompt that was actually used, which is the first thing to look at when a result surprises you.
• Content rules are applied at this stage. Refusals, and silent softening of a request, mostly happen in the rewriter.

Conversation as a dialect

Where the image model is itself a language model, or is tightly coupled to one, the prompt is a conversation, and the conversation can include pictures.

• Ask, then correct. “Make the lamp the only light source” and “move her to the left” are instructions a keyword model could never follow. Part 6 explained the editing models that make this work.
• Show it things. Upload a reference for the style, a photograph of the product, a sketch of the layout, and say what role each plays: “use the first image for the colour palette only”.
• Expect drift. Each edit regenerates the image, and small things change that you did not mention. After three or four rounds, details from the first version have usually gone. When a conversation has established what you want, write it out as one complete prompt and start again from that.
• It is slow and hard to repeat. There is rarely a seed to fix, and the model behind the service changes without notice. Part 6’s comparison of hosted and local applies in full.

Services with a house style

Some hosted services are tuned hard towards an attractive, recognisable look, and are controlled less by description than by a handful of switches. The names differ. The switches are nearly always the same.

On these services a short, evocative prompt often beats a long specification, because the house style fills the gaps attractively and a long prompt fights it. One well-known service, for example, takes its switches as flags typed after the prompt. When the house look is what you want, lean on it. When it is not, turn stylisation down before you write more words.

077. Motion, sound and shape

Video

Parts 3 and 6 recommend starting every shot from a still. When you do, the still has already answered most of the seven slots, and the prompt has a different job: it describes change. Repeating what the picture shows wastes the prompt and can fight the image.

A video brief has four slots of its own.

She spins the wheel slowly with her left hand and watches the rim pass the gauge. The camera dollies in slowly from a medium shot to a close-up of her hands and the rim. Steady, unhurried movement. The shot ends with the wheel still turning. Sound: the tick of the freewheel, a radio playing quietly in the background, rain on the window.

• One action and one camera move per shot. A sequence of actions is a sequence of shots, as part 3 explained.
• Use film language. Static shot, pan, tilt, dolly, tracking shot, orbit, crane, handheld, rack focus. These are the words in the training descriptions.
• Say what stays still. “The camera is locked off” and “she stays seated” prevent the default drift and wander.
• Prefer motion that footage is full of. Walking, turning, pouring, weather, traffic, cloth and hair. Part 3 listed what goes wrong: precise interactions, sequences, anything with an exact outcome.
• Dialogue goes in quotation marks, with who says it and how, on models that generate speech. Keep it to a short line per shot.
• Text-to-video, when you must, needs the full seven-slot brief and the four above, in prose, in that order. Video models are all prose models, since they arrived after the change in text encoders.

Music

A song generator takes two prompts, as part 4 explained, and they are written differently.

The style prompt is a brief in keywords or a short sentence. Its slots are genre and era, tempo, mood, instrumentation, the kind of voice, and production. It is a keyword dialect even on current services, since that is how music is catalogued and described.

Late-1970s British folk rock, 96 bpm, warm and unhurried. Fingerpicked acoustic guitar, upright bass, brushed drums, a little fiddle. Female alto lead vocal, close and dry, with two-part harmony on the choruses. Live-room sound, no synths.

The lyrics are performed and not interpreted, so they are written as lyrics, with the structure marked by tags on their own lines: intro, verse, chorus, bridge, outro, and instrumental breaks. The tags are the main control over a song’s shape.

• Name the instruments you want and the ones you do not. “No synths” and “no drums” work better here than negation does in images, because style prompts are short and services handle exclusions.
• Tempo and key are hints. Part 4 explained that the model has no notion of either. State them, then check.
• Write singable words. Regular line lengths, natural stresses, and no tongue-twisters or unusual names. Spell a name the way it sounds if it is being mispronounced.
• Keep it short. A three-minute song is about 200 to 300 words of lyric. More than that is crammed or dropped.
• Avoid naming artists. Describe the sound, for the reasons part 6 gave, which apply with more force in music. Many services block artist names in any case.

For sound effects and instrumental beds the brief is a plain description of the sound, its source, the space it is in and its length: “heavy rain on a tin roof, close, interior, steady, 60 seconds, loopable”.

3D

Part 5 explained that the real input to a 3D model is a picture. So prompting for 3D is prompting an image model for a very particular kind of image, and the slots are constraints and not creative choices.

A single vintage steel bicycle repair stand, whole object visible and centred, on a plain light grey background. Three-quarter view from slightly above. Soft, even studio lighting with no strong shadows or highlights. Matte painted steel, worn rubber clamp pads. Product reference photograph, sharp focus throughout.

• One object, whole, centred, nothing cropped.
• Plain background.
• Flat, even lighting. Anything dramatic is baked into the surface, as part 5 described.
• Three-quarter view, or a matched set of front, side and back views using part 6’s reference techniques.
• Characters in a neutral pose, arms away from the body.
• Describe materials plainly. “Brushed steel”, “matte rubber”, “oiled oak”. The painting stage turns these into material maps.
• Avoid what 3D cannot hold. Hair in wisps, foliage, wires, glass, smoke, and fine lettering.

Where a service takes text directly, write exactly this description and let it make the picture.

088. Iterating with purpose

The loop

Part 1 gave the two modes. Exploring means one prompt and many seeds, because the seed mostly decides composition. Refining means one seed and one change at a time, because only then is the difference between two images caused by what you changed.

Three habits turn that into a method.

• Judge batches. Generate four to eight images before you touch the prompt. If one in eight is right, the prompt is nearly right and you need more seeds. If none is, the prompt is wrong and more seeds will not help.
• Change one thing. Rewriting the whole prompt after each disappointment tells you nothing. If three changes are needed, make them in three steps and keep the ones that helped.
• Write it down. Keep the prompt, seed and settings with each image you might return to. Tools for open models do this in the file itself, as part 6 described.

Diagnosing a bad result

The last but one row is the one most often ignored. Once an image is ninety percent right, the prompt has done its job. Regenerating in the hope of a perfect draw wastes time that ten seconds of inpainting would save.

Using a language model to write prompts

Turning a brief into a fluent prompt in a particular dialect is a translation task, and language models are good at translation. It works well under three conditions.

• Tell it the target. “Write a prompt for an SDXL fine-tune: comma-separated keywords, under 60 tokens, most important first, plus a short negative prompt” produces something usable. “Write an image prompt” produces 200 words of purple prose suited to no model in particular.
• Give it the brief, not the wish. Hand over the seven slots. Left to invent them, it chooses the same defaults the image model would have.
• Watch for sameness. Language models have favourite words and favourite pictures. Prompts written by them converge on a look: cinematic, golden-hour, highly detailed. If every image your team makes starts to feel alike, this is the usual reason.

Prompts as shared tooling

For a team, the unit of reuse is not a prompt but a template: fixed text for the things that define the brand’s look, which are usually medium, lighting, palette and finish, and named slots for what varies. Store it with the model and version it was tuned for, the settings, any LoRA and its strength, and a few reference outputs that show what good looks like. When the model changes, regenerate the reference outputs and compare, which is a small version of the evaluation discipline in part 9 of the practical AI module.

A prompt tuned on one model rarely transfers unchanged, even within a dialect. Budget time to re-tune templates whenever the model behind them changes, including when a hosted service changes it for you.

099. The cheat sheet

The table is for still images. For the other media: video models read prose, and the prompt describes motion and camera over a still you supply. Song models read a keyword style brief plus tagged lyrics. 3D models read a plain, evenly lit picture of one object.

Whatever the model or the medium, four rules hold.

Write the brief before the prompt. Seven slots for a picture, and the extra ones for motion or sound. Know which ones you left to the model.

Match the dialect to the training. Look at the publisher’s examples. They show what the model was taught to read.

Say what is in the picture, not what is not. Negation fails in every dialect except a true negative prompt.

When words stop working, stop using words. Layout wants a sketch or a control image. A consistent character wants a reference or a LoRA. One wrong detail wants inpainting. A video shot wants a still to start from, and a 3D object wants a picture. All of those are in part 6, and knowing when to reach for them is the larger part of prompting skill.

The whiteboard version

A prompt is a brief: subject, action, setting, composition, lighting, medium, mood. The model fills every gap with its most ordinary option. How you write the brief depends on what the model was trained to read. Older models read comma-separated keywords inside a 75-token window, with bracket weights and a negative prompt. Tag-trained fine-tunes read their own tag vocabulary. Current models read a few plain sentences, most important first, and can follow “on her right” and lettering in quotation marks, but not negation. The newest read labelled fields or JSON, which helps your workflow more than your pictures. Assistants rewrite your prompt and let you correct the result in conversation. For video, describe what moves and what the camera does over a still you made first. For music, a keyword style brief and tagged lyrics. For 3D, a plain picture of one object. In every case you improve a prompt by judging batches, fixing the seed and changing one thing at a time, and you stop prompting and start editing once the image is nearly right.