AI prompting: your questions answered.

What is an AI prompt?

A prompt is the written instruction you give an AI tool. For pictures and films it is a description of exactly what the camera should see: who or what is in the shot, what they are doing, where, how it is lit and how it is filmed.

The tool can only show what you describe. Anything you leave out, it fills in by itself — usually not the way you pictured it. A good prompt closes those gaps.

What is the single most important rule?

Write what can be seen. AI models respond to things they can picture and measure, not to mood words like “epic”, “tense” or “cinematic”.

Turn every feeling into something visible. Instead of “a tense scene”, describe what tension looks like: a man freezes, slowly clenches his fist, light falls only from one side so half his face is in shadow.

Vague:   a cool cinematic shot of a car, epic, fast

Visible: low tracking shot alongside a black sports car as it powers through a wet bend at 90 km/h, headlights glowing, spray lifting off the rear tyres, the camera shaking slightly with the speed, cold blue dusk light

Should I say what I don’t want?

No — describe what you do want. “Not blurry” or “doesn’t fall over” often puts the very thing you are trying to avoid into the model’s head.

Swap every negative for a positive: “stays upright, feet planted” instead of “doesn’t fall”; “sharp focus on her eyes” instead of “not blurry”. Some image tools have a separate “negative prompt” box for unwanted things — use that box for them, and keep the main prompt positive.

How long should a prompt be?

As long as it needs to be to remove guesswork, and no longer. Fewer precise words beat lots of vague ones.

A quick image can be one strong sentence. A film shot with a camera move, a performance and a lighting change may need a structured prompt of 150–300 words. If a sentence doesn’t change what the camera sees, cut it.

What order should I write things in?

Subject first, then what they are doing, then where, then the camera, then the light, then the finish (style and format). The model pays most attention to what comes first.

A useful checklist: who or what → appearance (only what matters) → action → setting → camera and lens → lighting → colour → style. Leave out anything the shot doesn’t need.

A woman in her thirties with short silver hair and a camel wool coat, laughing as she turns towards the camera, standing on a wet London street at night, medium close-up at eye level, soft neon spill from a shop window on her left, warm 3200K street lamps behind her, shallow depth of field, photoreal, fine film grain

How much should I describe a person?

Only the details that matter to the shot: age, build, hair, one or two signature pieces of clothing, and anything the action depends on.

Long lists of features fight each other and the model drops some of them. If you are using a reference picture of the person, keep the written description very short — let the picture set the look and let the words set what happens.

How do I show emotion?

Describe the muscles, not the label. The model can draw a swallow or a tightened jaw; it can’t draw “sad”.

“Her eyes drop to the table, her jaw tightens, she swallows once before answering” works far better than “she looks sad”. Restraint usually reads better than big gestures.

How do I use reference images?

Upload the picture, then refer to it by its tag in the prompt — for example @image1 — and say what it is for: the character, the product, the location or the style.

Newer video models such as Seedance 2.0 take several references at once — up to nine images plus video and audio clips — and each is tagged so you can give it a role. Only mention a tag in a shot where that person or object actually appears, or the model will try to force it into the frame.

Small but important details — a logo, a number plate, the colour of a button — are worth writing in words as well, because models can drop fine detail from a reference.

@image1: man in his forties, broad build, grey wool overcoat, silver signet ring on his right hand. 100% matches the reference.

@image1 steps out of a black car onto a rain-soaked cobbled street, pauses, and looks up at a lit doorway on the left of frame. Medium shot at chest height, camera slowly pushing in. Cold 5600K moonlight from behind, warm amber light spilling from the doorway onto his face.

How do I write a good image prompt?

Name the subject, the moment, the setting, the framing and the light — in that order — then finish with the look you want.

A pair of clear glass perfume bottles on a slab of wet black slate, a single drop running down the front bottle, close-up from a low angle, hard white rim light from behind so the glass edges glow, soft grey fill from the front, fine mist in the air, deep black background, photoreal, crisp studio detail

How do I get the lighting right?

Say where the light comes from, how hard it is and what colour it is. Colour temperature in Kelvin is the clearest way to describe colour.

Roughly: 3200K is warm tungsten and candlelight, 4000K is neutral, 5600K is daylight, 8000K+ is cold blue shade or moonlight. “Hard light from the left, soft fill from the right, 3200K” tells the model far more than “moody lighting”. Tie colour to an object and a light: “a crimson silk scarf catching cold light from the corridor” beats “red and blue”.

What camera words actually work?

Shot size (close-up, medium, wide), camera height (eye level, low angle, overhead) and field of view. Describe the look, not brand names of cameras or lenses.

A rough guide to field of view: around 84° feels wide and spacious, 47° feels like the human eye, 29° is a flattering portrait, 18° or narrower compresses the background like a telephoto lens. Many tools understand both “85mm portrait lens” and degrees; degrees tend to be more consistent.

How do I keep the same character across pictures?

Use the tool’s character or reference feature rather than re-describing the person each time.

Higgsfield has Soul ID for a persistent character; Midjourney’s V8.2 edit model accepts up to 4 image references; most video models accept a reference image of the character. Keep the wording identical between prompts for anything that must not change — the coat, the hair, the ring.

How is a video prompt different from an image prompt?

You also have to describe movement and time: what moves, how fast, what the camera does, and how the shot begins and ends.

Describe the camera and the subject separately — “the camera tracks left at walking pace while she runs right” — and use real speeds (“40 km/h”) instead of “fast”. Say what is in the very first frame, so the shot doesn’t open empty.

How do I control cuts and timing?

Choose how much control you need: one continuous shot, a list of cuts in order, or cuts at exact seconds.

For a single take, say “one continuous shot, the camera does not cut on its own”. For a sequence, write CUT 1, CUT 2, CUT 3. When beats must land on the clock, give timecodes. Whichever you choose, keep the same people, wardrobe, light and direction of travel across every cut.

0.0s to 2.0s — wide shot, a lone rider on a red motorcycle crests a desert ridge at dawn, dust trailing behind at 60 km/h
2.0s HARD CUT
2.0s to 4.0s — close-up on the rider’s visor, the rising sun reflected in it, slight camera shake
4.0s HARD CUT
4.0s to 6.0s — low tracking shot beside the front wheel as it bites into the sand
Cuts only at the specified points. Same rider, same red bike, same low golden 4000K light in every shot.

What does a full film-quality prompt look like?

For our own films we write prompts in labelled blocks, so nothing is left to chance: context, references, first frame, camera, action, physics, lighting, then style and a few “locks” at the end.

SCENE CONTEXT
A vintage silver sports car arrives at a cliff-top villa at sunset.

FIRST FRAME
The car is entering frame from the right on the gravel drive; the villa’s glass front fills the background.

OPTICS
Wide 63° field of view, eye level.

CAMERA
One continuous shot, the camera does not cut on its own. Slow dolly left at 5 km/h, keeping the car centred.

ACTION
The car rolls to a stop at 10 km/h; the driver’s door opens.

PHYSICS
Gravel crunches and kicks up a little dust under the tyres; the car settles on its suspension.

LIGHTING
Low sun behind the camera, warm 3200K light across the paintwork, long shadows to the left.

STYLE
Photoreal, fine film grain.

LOCKS
Silver paintwork stays clean and reflective; headlights stay on.

Can AI video tools make sound?

Yes — several now generate sound with the picture. Seedance 2.0 and Google’s Veo 3.1 both create dialogue, sound effects and ambience together with the video.

Describe only the sounds you need: “rain on the car roof, distant thunder, no music” — or the exact line someone says, in quotes. If you plan to add your own music later, ask for ambience only.

What is Higgsfield?

Higgsfield is an AI video and image platform from San Francisco, founded in 2023, that puts many of the leading generation models in one place alongside its own tools.

Rather than one model, it gives you a choice — video models including Seedance, Kling and Veo 3.1, and image models including Nano Banana, Flux 2, GPT Image and its own Soul — plus studios built on top of them: Cinema Studio for camera, lens and lighting control, Soul ID for keeping a character consistent, Lipsync Studio, Marketing Studio for ads, and upscaling.

It works on credits, with monthly plans. It’s a good fit if you want to compare models side by side, or move a character from stills into film without switching apps.

Which video model should I use?

Start with the job: references and multi-shot storytelling, a single beautiful shot, or dialogue with sound — then test two models on the same prompt.

Seedance 2.0 (ByteDance, released February 2026) takes many tagged references and can plan several shots from one prompt, with sound. Kling 3.0 (Kuaishou, February 2026) is built to handle complex, fast motion. Veo 3.1 (Google) makes clips of up to 8 seconds with synchronised sound. OpenAI shut down its Sora video app in 2026.

Models update every few months, so the “best” one changes. Running the same prompt through two of them on a platform like Higgsfield is the quickest way to find out which suits your shot.

Which tools are best for still images?

Midjourney for art direction and style, Google’s Nano Banana and OpenAI’s GPT Image for following detailed instructions and editing, and Flux or Seedream for photoreal detail — most are available inside Higgsfield too.

Midjourney’s current model is V8.2 (July 2026), with style references (--sref), a draft mode that makes 24 quick variations, and an edit model that takes up to 4 reference images. Whatever the tool, the prompt principles on this page are the same.

Can ChatGPT or Claude write prompts for me?

Yes, and it’s one of the best uses for them. Describe the shot in plain English and ask for a prompt in a fixed structure — then edit it yourself.

Give the assistant your rules (subject first, positive wording, Kelvin, field of view in degrees, speeds in km/h) and a good example. You stay the director; the assistant does the typing.

Write a Seedance video prompt for this idea: a florist opening her shop at 6am, rain outside, first customer arrives. Use blocks: SCENE CONTEXT, FIRST FRAME, OPTICS, CAMERA, ACTION, LIGHTING, STYLE, LOCKS. Describe only what can be seen, positive wording only, colour temperature in Kelvin, field of view in degrees, one continuous shot.

Can you prompt a whole website?

You can prompt a very good starting point. Every one of our website builds began as a written brief, then was designed and hand-coded from there.

The trick is the same as with film: describe what the visitor sees and does, section by section, and the one interaction that makes it memorable. Name the audience and the single action you want them to take.

A one-page website for a boutique hotel in Marylebone, London. Audience: couples planning a weekend away. Goal: a reservation request. Sections in order: hero with an evening film of the facade; rooms with prices; a dial the visitor drags to change the facade lighting by hour; restaurant and bar; walking times to nearby sights; reservation form. Warm near-black, brass and rose; large serif headlines with one italic word. Phone first, fast loading, no stock-photo look.

Can I use AI images for my business?

Usually, but check the terms of the tool you used and the plan you’re on — commercial rights vary between tools and plans.

Don’t create images that pass off a real person, brand or product as endorsing you. If you make a concept for a real brand (as we do in some portfolio pieces), label it clearly as an unofficial concept.

Why does my AI video look wrong or warp?

Usually because something was left to guesswork — an empty first frame, a vague action, or too many things moving at once.

Fixes that work: describe the first frame exactly; give one main action per shot; separate camera movement from subject movement; add physics (“the car settles on its suspension”, “rain runs down her hair”); and finish with a short “locks” line restating what must not change.

Why does my character change between shots?

Each generation starts from a blank slate with no memory of the last one.

Write every prompt as if it stands alone: no “as before”, no scene numbers, no characters who aren’t in this shot. Use a reference image or a character feature like Soul ID, and repeat the key details word for word.

Why is the colour or mood off?

Colour was given as a list instead of being tied to light and objects.

Set one colour temperature for the scene in Kelvin, say where the light comes from, and attach each colour to a material: “teal light from the aquarium across her pale cheek” rather than “teal and pale”.