The 6-Slot Prompt Framework

Published October 2026 · 11 min read

Most bad AI video prompts fail for the same reason: they describe a mood instead of a scene. “Cinematic drone shot of a beautiful city at sunset, highly detailed, 4K” sounds like a prompt. It is actually a wish. It tells the model how you want to feel about the result, and almost nothing about what to render.

The fix is not a longer prompt. It is a structured one. After a few thousand generations across Seedance, Veo, Sora and Kling, we settled on six slots that, filled in order, cover everything a video model needs to make a decision. Fill them all and you rarely get a bad surprise. Leave one empty and the model invents something — usually the wrong thing.

Why order matters

Video models do not weight every word equally. Early tokens anchor the composition; later tokens adjust it. If you open with camera language (“aerial shot pushing in…”), the model commits to a framing before it knows what is in the frame. If you open with the subject, everything after it modifies a fixed anchor, which is far more stable.

So the slots run from content to form: what is there, what it is doing, where it is, how it is lit, how it is filmed, how it is rendered. That order mirrors how a director would brief a shot, and it happens to match how these models build an image.

The six slots

Slot 1 — Subject

Who or what is in frame, described with enough specificity to rule out the statistical average. Age, build, clothing, distinguishing features, and condition — weathered, tired, freshly showered, mud on the boots.

a 52-year-old fisherman, sun-damaged skin, grey stubble, oil-stained canvas jacket

The most common mistake here is using evaluative words instead of descriptive ones. Beautiful, handsome, professional and attractive all resolve to a retouched stock-photo look. Describe the features you actually want and the model will build a person rather than a category.

Slot 2 — Action

What the subject is doing at this moment. Not a state (“standing”) but a specific action with an implied direction and intent (“stepping off a curb into the road, glancing left”).

stepping off the curb into the road, glancing left, one hand raised to flag a taxi

Specific action is what turns a still image into a video. It also gives the model a reason to generate motion, which is the thing you are paying for. A prompt with no action often produces a slow drift or a static hold — technically video, creatively useless.

Slot 3 — Setting

Where and when. Location, time of day, weather, and one or two environmental details that establish scale or atmosphere.

rain-slicked side street in a port town, early morning, low fog, fishing boats moored behind him

Resist stacking adjectives here. Two or three concrete details do more than ten atmospheric ones. “Fog” and “wet asphalt” are things the model can render; “moody” and “atmospheric” are not.

Slot 4 — Light

The single most under-specified slot, and the one that most affects whether the result looks real. Name the source, the direction, the quality and the colour.

low overcast light from behind, soft rim on the shoulders, cool grey-blue cast, no hard shadows

Source words beat mood words every time. “Golden hour” is a time; “low warm sun from camera left, long shadows across the pavement” is a lighting setup. The lighting phrase library has source-based descriptions grouped by time of day and interior type.

Slot 5 — Camera

Shot size, angle, and movement — one movement, with a speed. Combinations are where things fall apart.

medium shot, slight low angle, slow steady push in, handheld with minimal drift

Shot size is doing more work than people realise. A close-up invites scrutiny of the face and eyes, where models are weakest. A medium or wide shot hides those weaknesses and reads as more cinematic. If a generation keeps failing, the cheapest fix is often to pull the camera back rather than rewrite the prompt. Our camera move reference lists which movements hold up reliably and which tend to break.

Slot 6 — Texture

The rendering character: film versus digital, grain, contrast, colour treatment, lens behaviour.

35mm film look, fine grain, slightly muted colours, shallow depth of field

This slot is optional and should stay short. Two or three texture terms are plenty; a pile of them starts to fight the rest of the prompt. It is also the slot most responsible for making AI video not look like AI video — a little grain and a slightly imperfect colour grade do more for believability than any amount of “8K ultra-detailed”.

The full prompt, assembled

Put the six slots together in order and you get this:

a 52-year-old fisherman, sun-damaged skin, grey stubble, oil-stained canvas jacket, stepping off the curb into the road, glancing left, one hand raised to flag a taxi, rain-slicked side street in a port town, early morning, low fog, fishing boats moored behind him, low overcast light from behind, soft rim on the shoulders, cool grey-blue cast, medium shot, slight low angle, slow steady push in, 35mm film look, fine grain, shallow depth of field

It is long, but every clause is load-bearing. Remove any one and the model has to invent that dimension. That is the whole argument for structure: length is a consequence of completeness, not a goal in itself.

Three failure patterns to avoid

Slot conflict. “Hard midday sun” in slot four and “soft diffused shadows” in slot six. The model averages them and you get flat, ambiguous light. If you notice a result that looks oddly generic, check whether two slots are contradicting each other.

Adjective stacking. “Beautiful stunning gorgeous breathtaking cinematic” is one word repeated five times with increasing desperation. Models do not sum synonyms into intensity; they dilute the signal. Pick one and spend the saved tokens on a concrete detail.

Empty action. A prompt that describes a person and a place but never says what is happening will produce something between a photograph and a slow zoom. Always give the model a verb.

Model differences worth knowing

The six slots work across models, but sensitivity differs. In our testing:

Model-specific guidance is in the individual guides: Veo, Kling, Sora, Seedance.

A second worked example: interior dialogue

Outdoor single-subject shots are forgiving. A two-person interior scene exposes every empty slot immediately, which makes it a good test of whether the framework actually holds up.

Slot 1 — Subject. Two people, described separately, because the model needs to keep them distinct across frames:

two people in their thirties facing each other across a small cafe table, one in a worn denim jacket leaning back, one in a grey knit sweater leaning forward

Slot 2 — Action. This is the slot that decides whether the shot works. If you write “having a conversation”, both mouths move at once and the result is uncanny. Name who is speaking:

mid-conversation, the person in denim talking with one hand raised, the other listening and nodding

Slot 3 — Setting. Keep it to three concrete details:

narrow corner cafe, late afternoon, half-empty, rain on the window behind them

Slot 4 — Light. Interiors need a named source more than exteriors do, because the model will otherwise light both faces evenly and flatten the scene:

soft window light from camera right, warm, falling across the near side of both faces

Slot 5 — Camera. A static two-shot removes an entire class of failure. Dialogue scenes do not need movement:

two-shot, eye level, static tripod, shallow focus favouring the speaker

Slot 6 — Texture. Short and neutral:

clean digital, natural colour, mild contrast

Note what changed between the two examples. The slots are identical; only the content differs. The framework is the same for a fisherman on a wet street and two people in a cafe — which is the point. Once the slots are habitual, you stop rewriting prompts from scratch and start filling in a form.

How long is too long

There is a practical ceiling. Beyond roughly 150–200 words, additional clauses start to dilute rather than refine — the model has a finite attention budget and every token competes for it. If your prompt is running past that, the problem is usually that you are trying to describe two shots in one. Split them and generate separately.

The prompt length checker flags when you are in the dilution zone for a given model.

Using the framework in practice

Two ways to apply this. The manual way: write your prompt in six lines, one per slot, then join them with commas. Writing in lines first makes empty slots obvious — a blank line is a gap the model will fill for you.

The faster way: the prompt generator presents these six fields in order and assembles the result, with model-specific presets for camera and lighting vocabulary. Either approach produces the same structure; the point is that the structure is the same every time.

Once the habit is in place, the framework stops being a checklist. You will find yourself thinking in six slots before you open the tool — subject, action, setting, light, camera, texture — and the retry count will drop because the model is no longer guessing.