AI Motion Graphics: Recreate Any Reference Video With Claude and Google Flow

By Clayton H · Published 7 November 2026 · Tools: Claude, Google Flow

You can make your own motion graphic from a reference video in three steps. Find a clip whose style you like, upload it to Claude with the reverse-engineering prompt below so it writes a detailed, shot-by-shot video prompt, then paste that prompt into Google Flow, choose Omni Flash as the video model and generate. You do not need After Effects or any animation skills.

What is reverse-engineering a video prompt?

Reverse-engineering means working backwards from a finished clip to the prompt that could produce it. Instead of guessing words like “cinematic” or “dynamic”, Claude watches the reference and describes what is actually on screen: the framing, how the camera moves, what changes and when, the light, the colour grade and the texture. A video model follows that kind of concrete, timed description far more reliably than a vague style name.

What you get

• The reverse-engineering prompt, ready to paste into Claude with your reference clip.
• The three-step workflow: find, break down, generate.
• Tips for picking references that AI video models can actually recreate.

What you need

• A reference video. A short clip (a few seconds) from Pinterest or anywhere you save inspiration. Download it so you can upload the file.
• Claude with file uploads.
• Google Flow with access to the Omni Flash video model.

How to make AI motion graphics in 3 steps

Step 1: Find a reference

Search Pinterest for the look you want (for example “motion graphics loop”, “abstract 3D animation” or “fashion motion design”) and save a short clip that matches the visual style you are going for. One clear idea per clip works best.

Step 2: Turn it into a prompt with Claude

1. Open a new chat in Claude.
2. Upload the reference video.
3. Paste the reverse-engineering prompt below and send it.
4. Claude measures the clip, builds a timed shot list and then writes one generation-ready video prompt. Copy only the final prompt.

Step 3: Generate in Google Flow

1. Open Google Flow and start a new video.
2. Choose Omni Flash as the video model.
3. Paste the prompt from Claude and hit Generate.
4. Compare the result with your reference side by side. If something is off, tell Claude what changed (for example “the camera moves too fast” or “the colours are too saturated”) and ask it to adjust only that part of the prompt.

The reverse-engineering prompt

You are a motion designer and video prompt engineer. I have uploaded a reference video. Your job is to reverse-engineer it into one prompt that an AI video model can use to recreate its look and motion as closely as possible, without copying any logos, text or recognisable characters.

Work in this order and do not skip steps:

1. Measure the clip: state its duration, aspect ratio and whether it loops (does the last frame match the first?).

2. Build a shot list: watch the whole clip and write a timecoded list. For each moment, note what is on screen, what moves, the direction and speed of the movement, and any cut or transition. Describe only what you can actually see.

3. Write the prompt in these labelled parts:

THE PIECE: what it is, in one sentence.

FORMAT: aspect ratio, duration and whether it should loop seamlessly.

OPENING FRAME: exactly what the first frame shows.

LAYOUT: where every element sits in the frame (front, middle, back, left, right, centre).

TIMELINE: what happens, in order, with timecodes.

CAMERA MOVE: the camera’s movement with timing, for example “slow push-in from 0 to 3 seconds, then still”.

LENS: lens feel, depth of field and focus.

MATERIALS AND MOTION: what things are made of and how they behave (weight, bounce, liquid, cloth, particles).

LIGHT: light sources, direction, softness and any changes over time.

COLOR: palette with hex codes, contrast and saturation.

LOOK: the visual style in concrete terms (shapes, surfaces, render look, texture, grain).

SOUND: sound design if relevant, or “no audio”.

MUST-KEEP: the five things that matter most to the look, written as positive instructions.

4. Give me the final prompt in one copyable block. Keep it specific and visual; avoid vague words like “cinematic”, “epic” or “dynamic” unless you also say what they mean on screen. Do not include text, logos or watermarks in the scene.

How to choose a good reference

• Short and simple. A few seconds with one main idea recreates better than a busy montage.
• Clear motion. Smooth, readable moves (a rotation, a push-in, a morph) are easier to describe than chaotic cuts.
• Few or no words. AI video models still struggle with readable text, so avoid references where typography is the whole point.
• Material you can name. Glass, chrome, liquid, fabric and paper give Claude concrete words to work with.

Things worth knowing

• It is a recreation, not a copy. The goal is the same style and motion with your own result. Do not repost someone else’s clip as your own, and credit the original creator if you show the two side by side.
• Expect a few runs. The first generation is rarely perfect. Small, specific fixes to the prompt work better than rewriting it from scratch.
• Timing drifts. Video models do not always hit exact timecodes. If a move comes too early or late, give it more time in the prompt rather than more words.
• Credits and limits depend on your plan. Check your Google Flow account for how many generations you have.

FAQ

Can AI make motion graphics without After Effects?

Yes. A video model such as Omni Flash in Google Flow can generate a motion graphic from a text prompt. The quality depends on how precisely the prompt describes the motion, which is why the reverse-engineering step matters.

Why use Claude to write the prompt?

Claude can watch the uploaded clip and describe it shot by shot, which gives you a far more specific prompt than you would write from memory.

Do I need a paid plan?

It depends on your Claude and Google Flow plans and their current limits, so check your own accounts.

Does this work with other video models?

The prompt is written in plain, labelled parts, so other video models that accept long text prompts can follow it. It was tested with Omni Flash in Google Flow, so test one run first if you switch models.

Can I use this for client work?

Use references for style, not as footage to copy, and check the terms of the video tool you generate with before using results commercially.