How to Write a Video Prompt Claude Can Follow

Why a video prompt needs more detail than a chat message

A video prompt has to describe what a viewer will see, because Claude cannot guess your taste. Anthropic’s prompting guide compares Claude to “a brilliant but new employee who lacks context on your norms and workflows”. The same guide asks you to be specific about the output format and constraints, and to write instructions as numbered steps when order matters.

It also suggests a quick test: show the prompt to a colleague who knows nothing about the project. If they would be confused, Claude will be too.

Motion makes that gap wider. “Smooth and modern” means something different to every designer. “The card rises 40 px over 0.4 s with easeOutCubic” means one thing. Anthropic’s guide also notes that animations should be asked for explicitly when you want them, rather than left for the model to infer.

Keep in mind who reads the prompt. Claude builds the video as code and renders it frame by frame, as explained in how AI renders motion graphics. Every value you leave out becomes a default someone else chose.

Start with the output spec and a timed scene list

The first lines should fix frame size, frame rate and duration. These three numbers decide everything else: where elements fit, how many frames each move gets and how long the render takes.

Output: 1080×1920 (9:16), 30 fps, 12 s, MP4. Silent; I will add music later.

At 30 fps, a 0.4-second move is 12 frames. That is worth knowing when you later ask for something “a little faster”.

Then list the scenes. Give each one a start time, an end time and one line on what the viewer sees.

Scene 1 (0.0–2.0 s): Dark background. The app icon scales from 0 to 100% at the centre.
Scene 2 (2.0–6.0 s): A phone slides up from below the frame and stops at the centre.
Scene 3 (6.0–9.5 s): Three feature cards stack to the right of the phone.
Scene 4 (9.5–12.0 s): Logo and a one-line tagline. Hold the last frame for 1 s.

Timed scenes do two jobs. Claude can plan the whole timeline before writing code, and you can later point at a moment (“scene 3, at 7.2 s”) instead of describing it. If the scene times do not add up to the stated duration, you catch the mistake before anything renders.

Describe motion, layers, colour and camera with numbers

Each moving element needs a start state, an end state, a duration and an easing curve.

Motion and easing

Easing is the speed curve of a move. easings.net describes easing functions as specifying how fast a value changes over time, and lists them by name: easeOutCubic, easeInOutSine, easeOutBack and more. CSS has its own keywords. MDN describes ease-out as a sharp start that gradually loses speed, and cubic-bezier() lets you give an exact custom curve. Either vocabulary is precise; pick one and stay with it.

The phone rises 600 px from below the frame in 0.7 s, easeOutCubic.
The three cards enter 0.12 s apart. Each fades from 0 to 100% opacity
and rises 40 px over 0.4 s, easeOutBack.

The 0.12 s gap is the stagger. Without it, “one after another” could mean a quick ripple or a slow slideshow, and Claude has to choose.

Layering

Say what sits in front of what, from back to front. Code draws layers in order, so this line becomes the draw order directly.

Layers, back to front: background gradient → soft glow → phone → cards → text.
Card 2 slides behind the phone; card 3 passes in front of it.

Colour and type

Use hex codes, font names, weights and sizes. “Brand purple” is a guess; #7C5CFF is not.

Background #0E0F13, accent #7C5CFF, text #F5F5F7.
Headlines: Inter Bold, 96 px, max two lines. Body: Inter Medium, 40 px.

If the font is not installed on your computer, say where the font file is.

Camera

A code-built video has no physical camera. A camera move is a transform applied to the whole scene, so describe it like any other move.

Camera: push in from 100% to 108% across scene 2, easeInOutSine. No shake.
Scene 3: pan 200 px left to reveal the cards.

Vague versus precise phrasing, element by element

The table below shows the same intention written two ways.

Examples help too. Anthropic’s guide ranks them among the most dependable ways to shape the format and tone of what Claude produces. For video, the example is often an image. Claude Code lets you paste or drag screenshots into the prompt, so a still from a video you like, with one line on what to copy (“match the spacing, not the colours”), adds context that words alone miss.

If your prompt mixes brand notes, a scene list and reference descriptions, separate them. Anthropic recommends XML tags for prompts that combine instructions, context and examples; clear headings work for shorter ones.

Ask for a preview, then give fix notes one at a time

End every video prompt with a preview instruction, so a wrong guess costs seconds rather than a full render.

Before the final render, render scenes 1–2 at 540×960 and export a contact
sheet of six frames from scene 3. Wait for my approval.

Half the width and half the height means a quarter of the pixels per frame, which is enough to judge timing and order. The contact sheet puts you and Claude in front of the same frames. Claude Code’s best-practice guide makes the same point in general terms: let Claude verify its work against something concrete, such as a screenshot set beside a reference, and ask to see that proof instead of a simple “done”.

Then fix one thing per message. Name the scene, the element and the second, say what should change, and close the door on everything else:

Scene 3, card 2: it arrives 0.2 s late. Start it at 6.3 s instead of 6.5 s.
Change nothing else.

One note per message keeps cause and effect clear. If three changes go in together and the result gets worse, you cannot tell which one caused it. Claude Code’s guidance adds a limit: correct early, and if the same issue still fails after two corrections, clear the session and start again with a better prompt. In practice, fold your fixes back into the scene list so the next run starts closer.

The free Combine & Fix workflow has ready-made notes for timing, motion feel, mask edges and render quality.

Put the blocks together in a fixed order

A complete prompt stacks these blocks in the same order every time: output spec, scene list, motion per scene, layers, colour and type, camera, then the preview instruction. A fixed order helps you spot a missing block, and it helps Claude find each value.

For a short real example, read the Orbit effect prompt: it gives radii as percentages, a revolution time, an entry easing and a front/back layer rule in about a dozen lines. If you have not prepared a project folder yet, the one-time setup prompt installs ffmpeg and the helpers those effects use.

Longer, scene-by-scene prompts for short product films are coming to the motion design prompts section for members. They follow the same anatomy, written out at the length a full product film needs.