Video prompt — what happens when you put six cuts into five seconds

What makes this hard

Video prompt failures are usually one of three. First, too many cuts. A cut needs at least 2 seconds to register, 3 with motion in it. Six cuts in a five-second video is under a second each, so no shot lands. Five seconds holds 1–2 cuts; thirty seconds, 6–10.

Second, the person changes when the scene changes. Text description alone will not bring the same face back. Third, another tool's syntax leaks in. A prompt carrying some service's name or flags like --ar gets refused by other tools. Or worse, the string gets read as picture content.

What to add to the request

1. Give total length; derive the cut count from it

Have the cut table carry real seconds as numbers and check that they sum to the total. Forbid 'divide length evenly by cuts' — different cuts need different time.

2. Fix one character phrase and repeat it in every scene

Settle the character's appearance in one phrase and repeat it verbatim in every scene prompt. Background and props not in the anchor will differ scene to scene, so settle them in sentence form in scene 1 and have later scenes cite it. If the tool takes reference images, also require the step of feeding scene 1's result as reference.

3. Parameters as plain values only

• Aspect ratio as '16:9', resolution as '720p' — values, not any tool's flag syntax
• Quality and length caps differ by service plan, so require the line 'if unavailable, step down one'
• Past ~5 seconds, require the method: generate cuts separately and join them

The five parts of the finished prompt

A video prompt visibly separates what gets pasted into the engine from what a person reads.

• Cut table — cut number, frame (as a noun phrase), camera, audio, length in real seconds. The cut count comes from the total length, and the lengths have to add up to it
• Per-cut generation prompt — one block per cut to copy and paste as is. The character anchor phrase and the continuity sentence settled in scene one appear in every cut. No placeholders are left in
• Exclusions (negatives) — what must not appear, as a list of noun phrases
• Parameters — aspect ratio, resolution, frame rate, total length. Values only, no tool names, with a line noting that a service plan may not offer the value
• How to make it — how to generate cuts separately and join them, the order that keeps continuity, and what was changed from your version and why

Direction changes; facts don't

This service's video instruction has the target AI direct, not assemble. If moving what you wrote into boxes were the whole job, you could do it yourself. How many scenes to cut, what to show first, where to cut away: those are the AI's call. Where it sees a better version it changes yours and says in one line what it changed.

Aspect ratio and frame rate are set by the medium, so instead of asking, it recommends a value with a one-line reason, such as 9:16 as the vertical short-form standard, with a note to say if the release is horizontal. Backgrounds, props, and lighting detail not covered by the anchor are never left as blanks: it decides them now and writes them as sentences, because a blank goes into the generator unfilled.

Checklist for writing it yourself

1. 1Did you give total length in seconds — not a cut count in its place
2. 2Did you give fixed descriptions for people and objects
3. 3Did you list what persists across scenes (wardrobe, lighting, grade)
4. 4Did you leave a slot to check licensing for music, fonts, and source footage
5. 5Are tool names and flags absent from the prompt

The Video studio takes three lanes: a video to generate now, a storyboard organizing the flow, and a character sheet building a reference set for one person. If the same person appears across scenes, the character sheet comes first.