Prompting Video Editor Content
Prompting for Video Editors
Given a script line and style direction, produce one shot entry in this format:
SHOT 03
Script line: "The coffee brews slowly as steam rises in the morning light."
Image prompt: Close-up of a ceramic mug on a wooden table, steam curling upward,
golden morning light through a window, warm color palette, shallow depth of field,
cinematic realism, 35mm lens look
Motion prompt: Steam rises and drifts slowly to the left; light steam wisp movement only;
camera holds a slow, subtle push-in; mug and table remain static
Consistency block: [paste locked style/character description here]
Notes: 3s clip, loop-friendly, avoid text/hands in frame
Never generate prompts without first collecting script, style, and technical specs (see Input below).
Progress:
- Step 1: Collect inputs (script, style, character/reference, technical specs, constraints)
- Step 2: Break script into shots (one action per shot, matched to voiceover timing)
- Step 3: Build a locked character/style consistency block
- Step 4: Write image prompt per shot (subject, action, setting, style, camera, lighting, mood)
- Step 5: Write motion/I2V prompt per shot (what moves, how fast, camera movement, what's static)
- Step 6: Check against target model's known limits (clip length, artifacts, risky terms)
- Step 7: Review shot-to-shot flow against the brief
- Step 8: Output shot list with editing notes
Step 1: Input Checklist
If any of these are missing, ask once, then state assumptions and proceed:
- Script/voiceover: exact lines, split by phrase or beat
- Subject: what it is, who it's for, key message
- Style: art style, color palette, mood (e.g. "3D cartoon," "cinematic realism," "crochet texture")
- Character/environment: locked description or reference images, reused every shot
- Technical specs: target model (Midjourney, Runway, Kling, Veo, etc.), aspect ratio, clip length limit, total runtime
- Constraints: banned terms, brand rules, known model weaknesses to avoid (hands, on-screen text, complex physics)
Step 4: Image Prompt Formula
[Subject] + [Action] + [Setting] + [Style] + [Camera: shot type/angle/lens] + [Lighting] + [Mood/color]
Use concrete, specific wording. Avoid vague adjectives like "nice" or "beautiful" — replace with sensory, technical detail ("warm rim light," "shallow depth of field").
Step 5: Motion Prompt Formula
State explicitly: what moves, how much/fast, what camera does, what stays static. Ambiguity here causes unwanted motion artifacts. Keep it to one primary motion + one camera move per shot.
Example 1: Input: Script line "A pair of sneakers spins on a pedestal, product launch style." Style: cinematic product photography. Model: Runway. Output:
Image prompt: White sneaker on a rotating dark pedestal, studio softbox lighting,
black gradient background, high-gloss reflective surface, product photography style,
85mm lens, shallow depth of field
Motion prompt: Pedestal rotates 360 degrees slowly and smoothly; camera locked static;
no other movement; lighting remains constant
Example 2 (bad vs. good): Bad: "A cool superhero flying in the sky looking epic" Good: "Male superhero in red suit, arms forward, flying through cloud layer at high altitude, motion blur on cape, sunlight rim lighting from behind, low-angle wide shot, cinematic color grade"
- Reuse one exact character/style block verbatim in every shot's prompt — this is the single biggest lever for visual consistency.
- Change one variable at a time when iterating on a prompt that's close but not right.
- Keep a running library of prompts that worked well for the target model/style combo.
- Match cuts to voiceover/music beats before generating — pacing decisions belong in the shot list, not fixed later in editing.
- Plan post-production fixes (color grade, transitions, speed ramps) as part of the output notes, since they hide common AI generation weaknesses.
- Don't skip the consistency block — inconsistent characters/environments across shots is the most common failure mode.
- Don't cram every shot's motion and camera move into one sentence with no priority — the model will drop or blend details.
- Don't ignore per-model limits (max clip length, known artifact types) — check before writing, not after generating.
- Don't use vague mood words alone ("epic," "beautiful," "amazing") without concrete visual descriptors backing them.
- Don't output prompts without stating the target model/tool — prompt syntax and strengths vary significantly between them.