The Master Guide / 06 Anchors and stills · Still prompts, storyboard frames and diagrams
The still-prompt structure
One order for every generated frame, plate and internal still. Every item is required, written in each model's own form; an item may be one short clause, but never left out (decided).
One order for every generated frame, plate and internal still. Every item is required, written in each model's own form; an item may be one short clause, but never left out (decided). Character, state and prop sheets are the exception: they stay flat, with subject, layout, palette lock and flat light only (see Photoreal anchors (Nano Banana Pro) → Character sheets; Stylised films (Seedream 5.0 Pro) → Turnaround, expression map, state sheets).
# | Block | What to write |
0 | Context | The story intent, the emotional beat and the goal of the shot: which moment of which scene, what the viewer should feel, what this frame must make clear |
1 | Subject | Who is there and their state at this instant: the pose from the ground up (weight, hips, shoulders, head, each visible hand, gaze), the expression held, what each hand holds. Once someone is cast, their identity role line and a short identity lock replace any description of the face: the structures that must match (eyes, brows, nose, mouth, jaw, hairline, marks), and no reshaping, slimming, ageing or beautifying |
2 | Environment | The place, its landmarks and dressing, the time of day |
3 | Camera placement | Height, distance, angle with tilt and roll, focal length or field of view, aperture, what is sharp. The lens and distance come from the relationship the shot plays [proven] (see Camera and lens → Distance from the relationship, then the lens); no very wide lens within about a metre of a hero or beauty face unless the treatment wants the distortion |
4 | Composition | The frame inventory, foreground to midground to background; where the eye lands; edge control: what each edge cuts, nothing important touching an edge, no limb cut at a joint |
5 | Lighting | One motivated source: its direction, quality and colour temperature, the practicals, where shadows fall and how dense they are; a ratio (2:1 gentle, 4:1 dramatic) where it helps. More under the table |
6 | Palette | A named palette tied to material and light: each leading colour with its material, the light on it and where it sits, plus the colours that must not appear. The 60/30/10 split with placement only when you or the brief ask for it |
7 | Surface realism | Human imperfections (the skin conditions of this life and this day, a crease, a stray hair) and the place's (wear, dust, a chipped edge), natural and fine, never a recipe of dirt, and never on the client's product (see Making it feel real → Skin, make-up and wear with a history) |
8 | Atmosphere | Haze, steam, rain or dust, only if it is really in the scene |
9 | Processing | The film's look block: grain, highlight roll-off, contrast, grade (see Still prompts, storyboard frames and diagrams → Text, sourcing and the look block) |
10 | Delivery | Aspect ratio and size |
Light, in more detail
A scene's first frame shows or places the key [proven].
Later frames may go contrasty, with an edge on the face [proven].
Use negative fill where shape matters [proven].
Skin, fabric, glass and metal each answer the source in their own way [proven].
A speaking face keeps the mouth lit, and a product keeps its label lit [proven].
The light's home is Light and look → One motivated, named source and Light and look → Light for a moving camera.
Grammar and bias
Still-frame grammar only. Describe the instant, never the motion: "her hand halfway to the latch", "caught mid-stride", "the cup tilted at the lip"; never "she reaches", "he walks" or "the camera pushes". Motion verbs smear poses.
Positive bias. Write what is there ("an empty corridor"), not what isn't.
Surgical negatives. One short line at the end, only for artefacts that keep appearing (a second person, a face in a mirror, stray letters), named as instances.
No quality tags. Never "hyper-realistic", "ultra-realistic", "8K", "HDR", "cinematic" or "masterpiece". Realism comes from the light, the lens and the imperfections.
Revision levers. Beside each prompt, in your take log (see Planning, budgets and takes → Your take log) and never sent to the model, note the two or three phrases you would change for the likely faults ("too flat → ratio to 4:1"; "too symmetrical → move the table to the left third"). Change one lever per attempt. Plates, sheets and internal frames get levers too.
No grids of composition options. A grid of variants in one image rests on weak evidence (one contest winner), and its layout leaks into later frames. To explore, write two or three placements in words, choose one, and generate it.
Each model's form [official guidance]. Where this Guide's order and a model's form differ, the model's form wins.
Nano Banana Pro (photoreal) | Seedream 5.0 Pro (stylised) | |
Form | Flowing prose in the order above; the context block becomes the opening sentences; no capitalised headers | Plain, compact sentences in the order above, the result and the style sentence first; no capitalised headers |
References | Ownership sentences first, one per image, in tag order | One sentence per image, naming it by what it shows, never by number |
Camera words | A real lens description | Cues in the drawn medium's terms, not physics |
Length | Under about 2,000 characters [best available, not yet validated] | Under about 2,000 characters [best available, not yet validated]; the maker advises under 600 words [official guidance] |
GPT Image's diagrams and cut-outs show no light, palette or atmosphere. They carry only the items that apply: subject, view, composition and edges, delivery (see Still prompts, storyboard frames and diagrams → GPT Image's two jobs and its version rule).