Direction
Turning images and a script into a scene prompt.
- From script to prompt
- Reference roles
- Rendering style
- Staging and performance
- Camera placement and movement
- Timed shots and exact dialogue
- Script & directorYou
- Co-directorClaude Opus, ChatGPT Astra
From script to prompt
Creator’s process- Bring in the director.Codex received the character, background and script, and turned them into detailed scene direction.
- Use a structured generation prompt.Higgsfield’s examples showed how to describe scene events and camera work. Their workflows use Blender blocking; the prompt structure was the useful part, and it works with direction written in plain text.
- Fill out the scene prompt.ChatGPT received the prompt structure, character, background, script and camera direction, and filled out the prompt for the scene.
In Higgsfield’s 3D Jutsu walkthrough and @adilinthewild’s Blender + Higgsfield article, camera and staging come from a 3D blockout rendered as a reference video. This process keeps their labelled sections and writes framing, camera paths, positions and eyelines straight into the prompt. Both sources are credited on Resources.
The structure has eight parts. The sections below cover the five that need the most thought.
| 1 | Part, duration, aspect ratio | One sentence on the scene, the visual era, and whether it stays in one place. |
|---|---|---|
| 2 | Reference roles | Which image controls appearance, identity, location and props. |
| 3 | Rendering style | How surfaces, hair, light and reflections are built. |
| 4 | World and staging | The place, where everyone stands, what they touch, where they look. |
| 5 | Animation and performance | Delivery, plus body, hand, face and background movement. |
| 6 | Camera language | What the camera does and why. |
| 7 | Timed shots and exact dialogue | Shot by shot: framing, one move or hold, action, the exact line. |
| 8 | Sound and essential continuity | Only what isn’t already established. |
Reference roles
Prompt frameworkEach attached image gets one job, limited to what it shows.
- Finished game-style scene image — primary appearance across the whole frame and every cut: characters, architecture, foliage, pavement, background people, light, reflections.
- Character image — identity: face, hair silhouette, proportions, wardrobe, accessories.
- Photographic environment — architecture, layout and colours, translated into the game style.
- Prop image — silhouette, colours, construction, material, markings.
- Voice sample — timbre, accent and delivery only.
List only references actually attached. When a scene is split into separate generations, repeat the roles in each.
GAME-STYLE SCENE IMAGE — PRIMARY APPEARANCE REFERENCE: Match this image's rendering across the entire frame: character, architecture, foliage, pavement, background people, lighting, and reflections. Preserve its simplified geometry and visibly textured game surfaces throughout every camera cut. Animate characters with deliberate pose transitions, restrained facial rigs, simple hand shapes, and repeating background walk cycles.
Use it when a finished game-style scene image is attached.
Rendering style
Prompt framework“PS3 style” alone leaves too much to guess. Describe what the render is made of, then describe movement separately.
Rendering: how it looks
The target look was a 2009 cutscene, PS3/Xbox 360-era in-engine render: visibly modelled silhouettes, angular clothing folds, hair in clumps or textured cards, painted surface detail, simple skin shading, restrained highlights, baked-looking light, soft shadow maps, modest texture resolution and broad, simplified wet reflections.
Animation: how it moves
Deliberate pose transitions, slightly rigid torso movement, simple weight shifts, short transitions between holds, simple hand shapes, restrained mouth and jaw movement, occasional blinks, and eyes that hold a target before changing. Background walkers use recognisable walk loops with grounded feet.
One exclusion sentence
Name drift that has actually shown up: photographic pores and hair strands, fluid motion-capture acting, constant facial micro-expressions. Skip film grain, VHS effects and pixel art; they don’t produce a period 3D render.
Staging and performance
Prompt frameworkThe scenes that worked best felt like real places: believable character placement and a background with life. Write those details into the prompt.
- Position and support. Where the character stands, which third of the frame, and what they lean on or touch.
- Screen sides and eyelines. Keep them consistent across cuts. If characters switch sides, show it.
- Prop custody. Who holds what, and how. A handoff is released only after the other hand grips it.
- Background life with a purpose. A worker sets down a carton and rolls it away; someone places the check and leaves. Keep the hero’s sightline clear, and leave private scenes empty.
- Performance. Delivery plus a purposeful gesture where it helps, then a settle. Stillness can carry a line.
The main character stands sideways to the railing, torso perpendicular to it, one forearm resting on the top rail with clear physical contact. Keep him predominantly on the right third, looking toward the active camera. Preserve his position and screen direction across cuts.
Contact, torso direction, frame placement and gaze in three sentences.
Camera placement and movement
Prompt frameworkWithout a Blender blockout, the prompt describes the camera itself. Pick each move for its effect: close shots for intimacy, wides for context, locked close-ups to land a punchline or realisation.
A complete shot instruction
Framing + physical action + target + endpoint.
Low beside her bag, frame its handles resting over her forearm. Hold still; rack focus once from the dragon detail to the main character's unobstructed face, then hold his recognition.
Terms that get mixed up
- Pan rotates the camera; truck moves it sideways.
- Tilt rotates up or down; pedestal raises or lowers the whole camera.
- Zoom changes the lens; dolly moves the camera closer or further.
- Rack focus changes focus only. The camera can stay still.
- Arcs and orbits need a start position and a path.
Use one dominant move per shot unless the combination is the technique, as in a dolly zoom. Check it’s physically possible: an actor can’t hold eye contact with a camera behind their head. More terms in the camera vocabulary.
Timed shots and exact dialogue
Prompt frameworkThe timed body is the shot list: time range, framing, one dominant move or hold, action and exact line for each shot. Silent moments still need picture.
4.19–9.10 — Hard cut wide, full body. Reveal his relaxed rail lean and the plaza. Almost imperceptible handheld sway; one small explaining gesture: “Imagine being so ridiculous, your ‘foolery’ gets credited for generations.” 9.10–10.16 — Hard cut to a locked face close-up. A brief beat, then absolute deadpan: “I can.”
Seconds and hundredths. The full prompt is under Worked examples.
Checks before you generate
- State the timecode format. At 30 fps,
MM:SS:FFframes run 00–29;4.19 secondsis decimal. - Each separate generation has its own timeline starting at zero.
- Look for gaps, overlaps and backwards cues. Make sure the action fits the time given.
- Never quietly shorten dialogue or move a locked cut to make it fit.
- When splitting into parts, cut at a meaningful beat and describe the final composition and prop state at the boundary.
Use the attached character, background and script to write a cutscene prompt. Structure it as: 1. Part, duration and aspect ratio 2. Reference roles 3. Rendering style 4. World and staging 5. Character animation and performance 6. Camera language 7. Timed shots and exact dialogue 8. Sound and essential continuity There is no Blender reference video: write camera movement, framing, character positions and eyelines directly in the prompt. Keep my dialogue and pauses exactly as written. Put any timing questions outside the copy-ready prompt.
A template written for this guide. The full direction instructions are in the downloadable skill on Resources.