How to write prompts

Five parts, in this order.

1. Style, shot type, who is in the frame.
2. What moves in the surroundings.
3. What the character does, phase by phase.
4. What must not move.
5. The quality tail.

2D-animated, medium close-up of the red-haired girl in a black jacket. Rain falls steadily behind her. She raises the cup with her right hand, takes one sip, then lowers it. Her left hand stays at her side, her feet stay planted, the background stays exactly as it is in the first frame. The camera stays completely still, no zoom. Smooth anime animation, natural motion, detailed facial expression.

It works better than free text because it clearly separates the scene, sound, and music.

Anchoring a picture as the first frame

Start the first field with the anchor:

Picture 1 is fully referenced at 0.00 seconds and belongs to [Shot 1]. The style, subject, colors and spatial relationships of Picture 1 stay consistent.

Name the free hand, the feet, the tail, the background - exactly the parts that are in your frame and take no part in the action.

For the background one exact phrase works: the background stays exactly as it is in the first frame.

Here the rule is the opposite. Pin down nothing.

If you write that hands, hair or the pose stay in place during sex, running or dancing, the motion will stop. Describe continuity instead: the motion is continuous and repetitive.

Speech goes into the first field, not the sound one.

A woman in her 30s (S2), on-screen, mid-pitch voice, says with quiet anger <d>[Russian] Ты обещал пятнадцатого.</d>

• the identity, the action and the delivery stay outside the tags;
• inside <d> there is only the language and the exact words;
• for a voiceover write says in an off-screen voiceover;
• a line across a cut: the <scenetrans> tag at both sides of the join;
• a line cut off by the end of the clip: the <cutoff> tag.

The first time a character appears, give the approximate age, whether they are on-screen or off, the pitch of the voice, the speech rate and the accent. The same number (S1) holds across all scenes. Silent characters need no numbers.

The first shot [Shot 1] carries no timecode. After that the timecodes go strictly in ascending order:

At 00:03.000, the camera cuts to [Shot 2] ...

Wordings for a cut: cuts to, transitions to, switches to. Dissolves and fades are written separately and only if they are really needed.

A change of framing without a cut is camera movement, not a new shot.

Two fields, and they must not be mixed up.

Write sound as sentences with a verb and a surface, not as a comma-separated list.

Rain taps against the window while traffic hums underneath

rain, traffic, footsteps

It is written straight into the description; there is no separate field.

Words in double quotation marks are drawn by the model as a caption. Write them in the language they should appear in, and do not use quotation marks for anything else.

Write plainly: the model understands euphemisms worse than direct words.

• state the position of the participants and what exactly happens, not the general mood;
• say where the focus stays: the focus stays on the thrusting;
• describe the sound separately and concretely - it is half of the impression;
• do not pin down body parts, see the section on motion.

• do not retell what the picture already shows: hair colour, clothes and surroundings are set by the frame, and extra words take the model’s attention away from the action;
• do not write empty labels like “beautiful”, “high quality”, “cinematic” - they describe no frame at all;
• adverbs work better than adjectives: slowly raises is clearer than slow movement;
• put the quality tail at the end, not at the start - at the start it takes room away from the action.

A calm scene without speech

integrated_multimodal_description: [Shot 1] Picture 1 is fully referenced at 0.00 seconds and belongs to [Shot 1]. 2D-animated, medium close-up of the girl at the window. Rain runs down the glass behind her. She lifts the cup with her right hand, takes one sip, lowers it. Her left hand stays at her side, her feet stay planted, the background stays exactly as it is in the first frame. The camera stays completely still, no zoom. Smooth anime animation, natural motion, detailed facial expression.

overall_soundscape: Rain taps steadily against the glass while a low room tone continues underneath. The cup clicks softly against the saucer.

non_diegetic_music: N/A

A scene with speech and a cut

integrated_multimodal_description: [Shot 1] Live-action, medium shot of a man in his 40s (S1), on-screen, low-pitch voice, standing in a kitchen. He turns to the door and says with tired calm <d>[Russian] Я же просил не приходить.</d> At 00:03.500, the camera cuts to [Shot 2], a close-up of the woman in the doorway (S2), on-screen, mid-pitch voice. She answers quietly <d>[Russian] И всё-таки я здесь.</d> The camera stays still in both shots.

overall_soundscape: A refrigerator hums under the room tone. Fabric rustles as he turns, and the door creaks once.

non_diegetic_music: N/A

integrated_multimodal_description: [Shot 1] 2D-animated, medium shot of the two characters on the bed. The man drives the motion, moving with steady force. The motion is continuous and repetitive. No climax, no pose change, no new action, no cuts. The camera stays completely still, no zoom. Smooth anime animation, strong male-driven thrusting, passive body recoil, natural jiggle.

overall_soundscape: Damp skin slaps against skin on each thrust while the mattress creaks quietly underneath. She draws a short breath between strokes.

non_diegetic_music: N/A