Make a brand film from code with Claude Code, HyperFrames and video-use

A 90 to 120 second motion graphics brand film, rendered from code, with a fact-checked opening claim, a script you approved, scenes you saw as stills before any render, and a self-checked final mix. No video editor involved.

By the AY Automate team · Updated Nov 22, 2026 · 14 min read

Build time
About 3 days of planning: script, scenes, assembly and review
Level
Intermediate
Stack
Claude Code, HyperFrames, video-use, a voice model
Main risk
Unsourced claims and unapproved numbers on screen

Inside this playbook

  • 5prompts, templates and code to copy
  • 2diagrams and real screenshots
  • 7steps, start to finish
HyperFrames · Catalog
HyperFrames catalog page: Motion ready to install, with categories for 3D, carousels, text and captions, code, transitions, data and charts, scenes and demos, and 3D title blocks
The HyperFrames catalog: ready-made scenes, captions, charts and transitions your agent can install and adapt. Screenshot taken 2026-11-22.

In short

  • Give Claude Code one long prompt with five inputs: a reference film, a visual style by name, your brand files, your braindump and the proof you approve. Start from Walid's prompt below.
  • Pin every version. Write npx hyperframes@<version> with a version you checked, never @latest.
  • Keep four human gates: the script, one of three storyboards, three stills per scene, then the final film. No full render before you approve the stills.
  • Build each scene as its own HyperFrames composition in parallel sub-agents, then assemble with video-use: cuts on word boundaries, 30ms audio fades, loudness at -14 LUFS.
  • Make the agent check its own work: transcribe the final mix, review a frame every 7 seconds, fix what is broken, and show you the issue list.

Who it is for: Founders, marketers and content leads who want a motion graphics film for their company and are comfortable running Claude Code in a terminal. You need a reference film you like, brand files and a few numbers you are willing to show.

Overview

What you will have at the end

When AY Automate started working with its 60th client company, Walid did not hire an animator. He asked Claude Opus 5.5 to make a motion graphics video showing the growth, and posted what came back. The film was inspired by an a16z video. The whole thing came from one long prompt and a few tools that render video from code.

This playbook turns that prompt into a repeatable build. You will have a 90 to 120 second film, the HTML scenes it was made from, the script you approved and a record of the checks the agent ran on its own output. Because the scenes are code, changing a line of copy or a brand color means editing a file and rendering again.

Diagram

The whole system, with the human gates

  1. Your inputs

    reference film, style by name, brand files, braindump, approved proof

  2. Fact-check the opening claim

    real data, sources cited on screen

  3. Write the script

    you approve it

  4. Three storyboard directions

    you pick one

  5. Build scenes in parallel

    one HyperFrames composition per scene

  6. Three stills per scene

    you approve before any full render

  7. Voice with word timestamps

    every animation beat synced to the spoken word

  8. Assemble with video-use rules

    cuts, fades, loudness

  9. Self-check pass

    transcript, frame every 7 seconds, fixes

  10. Final MP4

    saved to your Downloads and opened

Trigger or inputModel stepPerson decidesYour systemsResult

The agent does the work between gates. You decide at each gate.

Background

Who this is for and what you need

This fits a company that wants a short explainer or announcement film and has something true to say: a milestone, a product, a claim worth backing with data. It also fits a team that already uses Claude Code and wants to see what it does outside a code editor. It fits less well if you want live footage of people, since the rules below keep faces out unless you provide them.

Before you start

  • Claude Code with a plan you already pay for. Check the Claude Code cost calculator if you want to estimate usage for a long session.
  • Node.js 22 or newer and FFmpeg. HyperFrames needs both.
  • HyperFrames. Open source under Apache 2.0, free to use from the CLI. The Studio app has a free plan, and Pro is $25 a month if you want its built-in video, image, voice and music models.
  • video-use. Open source. It needs ffmpeg and an ElevenLabs API key, because ElevenLabs Scribe gives the word-level timestamps.
  • A voice model. ElevenLabs, MiniMax or similar. Check current pricing on their sites.
  • A music model. Lyria or ElevenLabs Music. Check current pricing.
  • A transcription service. ElevenLabs Scribe or Deepgram, for the self-check and word timing. Check current pricing.
  • A reference film you like, as a URL the agent can download.
  • Brand files: logo, colors and fonts.
  • An approved proof list: the clients, numbers and quotes you are happy to show, and nothing else.

Step 1

Install the tools and pin the versions

HyperFrames turns HTML, CSS, JavaScript and media into deterministic MP4 videos. Deterministic matters here: the same composition renders the same film, so a fix on scene four does not quietly change scene two. You can use it from the CLI, from coding agents such as Claude Code, Codex and Cursor through skills, or as a rendering core. No React is required.

The skills are what make the agent good at it. They teach the production loop: plan the video, write valid HTML, wire seekable animations, add media, lint, preview and render. HyperFrames ships 9 skills. The main one, /hyperframes, carries the contract and the references for film, captions, overlays, motion and look.

Config

Setup commands

# 1. Prerequisites: Node.js 22+ and FFmpeg must already be installed
node --version
ffmpeg -version

# 2. HyperFrames skills for your coding agent (pick "Core Skills")
npx skills add heygen-com/hyperframes

# Or, for Claude Code, install the plugin
claude plugin marketplace add heygen-com/hyperframes
claude plugin install hyperframes@hyperframes

# Keep the skills in step with the CLI version you pinned
npx hyperframes skills update

# 3. CLI by hand. Pin the version you checked. Never @latest.
npx hyperframes@<pinned version> init my-film
npx hyperframes@<pinned version> preview
npx hyperframes@<pinned version> render

# 4. video-use: follow the install steps in its README
#    https://github.com/browser-use/video-use
#    It needs ffmpeg and an ElevenLabs API key (Scribe transcription).
#    Put the key in your environment, never in a prompt or a committed file.

Replace <pinned version> with a real version you checked. Install video-use from its README, not from a copy of these commands.

Two more things to know about video-use. It lets the model work with video without watching it. The first layer is a packed transcript of about 12KB (takes_packed.md). The second is a filmstrip and waveform image, pulled only at decision points. Its pipeline runs Transcribe, Pack, LLM reasons, EDL, Render, Self-eval, and it keeps session memory in project.md.

Step 2

Write the inputs brief

The prompt is only as good as what you feed it. Walid gives the agent five inputs and nothing vague. Fill them in before you paste anything into Claude Code.

  • Input 1

    Reference film

    A film to match in structure and pacing. The agent downloads it and analyzes the shot list, pacing and sound design before it writes a word.

  • Input 2

    Visual style, by name

    Name a style and list its traits. Walid's example is 3Blue1Brown: black background, thin white lines, serif type, blue, yellow and green accents, shapes that transform, one idea on screen at a time.

  • Input 3

    Brand

    Logo file, colors and fonts. Give files and hex values, not descriptions.

  • Input 4

    What we do, in your words

    A braindump. Messy is fine. Your phrasing is what keeps the script from sounding like every other company film.

  • Input 5

    Proof you can show

    Clients, numbers you approve and quotes. This is a closed list. Anything not on it stays off screen.

Naming a style is the strongest single instruction you can give. On 2026-09-21 Walid asked Claude Code to explain what Jev is in a simple way for LinkedIn. He wrote one prompt and said: use 3Blue1Brown style. That one phrase carried the background, the line weight, the type and the pacing. The post got 50 reactions, 9 comments and 2 reposts.

Template

Inputs brief

FILM BRIEF FOR [COMPANY]

Length: 90 to 120 seconds
Reference film: [URL to the film whose structure and pacing we want to match]
Visual style, by name: [e.g. 3Blue1Brown, and the traits you want: background, line weight, type, accent colors, how shapes move]

Brand files
- Logo: [path to the file]
- Colors: [hex values]
- Fonts: [font names and where the files are]

What we do, in my words (braindump, messy is fine):
[paste it here]

Proof you may show (approved list, nothing else goes on screen)
- Clients: [names I approve]
- Numbers: [each number, with where it comes from]
- Quotes: [quote, who said it, written permission yes/no]

Opening claim to fact-check: [the one claim the film opens on]

Tools I have keys for: [voice model, music model, transcription service]
Where to save the result: [folder]

Save this as a file in the project folder. The agent reads it first, and you can reuse it for the next film.

Step 3

Run the prompt, and read the script before anything else

This is the prompt Walid started with, word for word. Copy it, swap in your company and let it cook. Everything in square brackets is yours to fill from the brief in step 2.

Prompt

Walid's brand film prompt

You are directing a 90-120s brand film for [COMPANY], rendered from code. No video editor.
Setup (install only what's missing, and pin versions):
- video-use (github.com/browser-use/video-use) for cuts, audio mix, and word-level timing
- HyperFrames (npx hyperframes@<pinned version>) for every animated scene
- A TTS voice (ElevenLabs, MiniMax or similar), a music model (Lyria or ElevenLabs Music), and Deepgram or Scribe for transcripts
Inputs I'm giving you:
- Reference film to match in structure and pacing: [URL]. Download it, analyze the shot list, pacing, and sound design before writing anything.
- Visual style to match by name: [e.g. 3Blue1Brown: black background, thin white lines, serif type, blue/yellow/green accents, shapes that transform, one idea on screen at a time]
- Brand: [logo file, colors, fonts]
- What we do, in my words: [braindump]
- Proof I can show: [clients, numbers I approve, quotes]
Process (do not skip steps):
1. Research the claim the film opens on and fact-check it against real data. Cite sources on screen.
2. Write the script. Show it to me. Wait for approval.
3. Propose 3 storyboard directions. I pick one.
4. Build each scene as its own HyperFrames composition, in parallel sub-agents. Deliver 3 still frames per scene. No full render until I approve the stills.
5. Generate the voice with word timestamps and sync every animation beat to the spoken word.
6. Assemble with video-use rules: cut on word boundaries, 30ms audio fades, loudness at -14 LUFS.
7. Self-check before showing me: transcribe the final mix and confirm every line is audible, extract a frame every 7s and review it, and fix what's broken.
8. Save the result to my Downloads and open it.
Hard rules: no fake screenshots, no faces unless I provide them, no revenue numbers unless I approve them, and no claim on screen without a source.

Pin a real version in place of <pinned version> before you paste it.

Read the structure, because it explains why it works. The prompt tells the agent what to install and how, what you are giving it, and then a numbered process with an instruction not to skip steps. Two of those steps end in the words approval or pick, which tells the agent to stop. The last line sets hard rules. We cover each part in the steps below.

The agent opens with research. It checks the claim the film starts on against real data and plans to cite sources on screen. Then it writes the script and waits. This is the first gate, and it is the cheapest place to fix the film. A wrong sentence here costs you a minute. The same sentence after rendering costs a rebuilt scene.

Step 4

Approve the script, pick a storyboard and approve stills

The prompt builds in three stops before the final film, and you add a fourth when the film comes back. Each gate has one job. Tell your reviewers what to look at, so they do not give notes on things that are not ready yet.

Approval gates
  • 1. Script

    What the agent shows
    Full script with the sources for each claim
    What you check
    Every line sounds like you. Every claim has a source. Nothing on the screen is outside your proof list. Length fits 90 to 120 seconds when read aloud.
    What to say if it fails
    Cut or rewrite named lines. Ask for the source of any claim you do not recognize.
  • 2. Storyboard

    What the agent shows
    Three directions, each with the shot sequence and the look
    What you check
    Which direction matches the reference film's structure and pacing, and your named style.
    What to say if it fails
    Pick one. Ask for changes inside that direction, not a fourth direction.
  • 3. Stills

    What the agent shows
    Three still frames per scene, with a note on what each shows
    What you check
    Brand colors and fonts are right. Text is readable. One idea per scene. No fake screenshots, no faces you did not provide.
    What to say if it fails
    Name the scene and the frame. The agent revises that composition only.
  • 4. Final film

    What the agent shows
    The assembled MP4 plus the self-check issue list
    What you check
    Every line audible. Pacing matches the reference. The opening claim and sources read cleanly. The issue list is empty.
    What to say if it fails
    Give timestamps. The agent fixes, re-assembles and runs the self-check again.

The rule that saves the most time is the one in step 4 of the prompt: no full render until you approve the stills. Stills are cheap to produce and cheap to judge. A full render is neither. Three frames per scene (start, middle and end works well) show the layout, the type and how the idea moves, without a wait.

Pick a storyboard direction fast. The point of seeing three is that you react better to options than to a blank page. If none is right, say what the closest one gets wrong and ask the agent to revise that one.

Step 5

Build each scene as its own composition, in parallel

Once you pick a storyboard, the agent builds each scene as its own HyperFrames composition, in parallel sub-agents. One scene per composition keeps a fix local. If scene five has a typo, you rebuild scene five.

Claude Code makes the sub-agent briefs for you, but you should know what a good one contains. It names the scene and its script lines, sets the one idea on screen, repeats the style rules and the brand files, lists the proof allowed in that scene, and sets a stop line. The brief below is a template for that.

Prompt

Per-scene sub-agent brief

You build ONE scene of a brand film as ONE HyperFrames composition.

Scene number and name: {{scene_id}}, {{scene_name}}
What this scene says (script lines, with the time window): {{script_lines}}
The single idea on screen: {{one_idea}}
Visual style, by name: {{style_name}} and its rules: {{style_rules}}
Brand: {{logo_path}}, {{colors}}, {{fonts}}
Allowed proof for this scene: {{approved_proof}}

Rules:
- Use the HyperFrames skills. Write valid HTML, CSS and JavaScript with seekable animations only.
- One idea on screen at a time. Match the pacing of the reference film.
- Leave timing hooks for the voice: every animation beat has a named start time that I can move to a word timestamp later.
- Any figure, number or claim on screen needs a source line from the approved proof list. If it is not on the list, leave it out.
- No fake screenshots. No faces. No revenue numbers.
- Lint and preview the composition.
- Deliver exactly 3 still frames of this scene (start, middle, end) and a one-line note on what each shows.
- Do NOT run a full render. Stop after the stills.

Work only inside {{scene_folder}}. Do not edit other scenes.

Give each sub-agent its own folder. Two agents editing the same files is how scenes overwrite each other.

You do not have to start from nothing. The HyperFrames catalog lists scenes, captions, charts and transitions your agent can install and adapt. On 2026-11-22 the catalog page listed 448 items, with categories for 3D, carousels, text and captions, code, transitions, data and charts, and scenes and demos. A chart scene is a good example: install a chart from the catalog, then restyle it in your named style with your approved numbers.

Step 6

Generate the voice, sync the animation and assemble

After you approve the stills, the agent generates the voice with word timestamps and syncs every animation beat to the spoken word. This is why the scene brief asks for named start times on each beat: the agent moves them to the timestamps once the voice exists. A line of text that lands on the word it describes feels intentional. The same text a half second early reads as a mistake.

Then the agent assembles the film with video-use rules. Video-use cuts filler words and dead space, applies 30ms audio fades at every cut, and can generate animation overlays through HyperFrames, Remotion, Manim or PIL in parallel sub-agents. It also self-evaluates the rendered output at every cut boundary. You set the spec once in the prompt.

Audio and assembly spec
  • Cut points

    Value
    On word boundaries
    Why it matters
    A cut inside a word is audible as a click or a clipped syllable.
  • Audio fades

    Value
    30ms at every cut
    Why it matters
    A short fade removes the pop at each join without softening the voice.
  • Loudness

    Value
    -14 LUFS
    Why it matters
    A fixed loudness target keeps the film from sounding too quiet or too loud when it plays next to other content.
  • Timing source

    Value
    Word-level timestamps from the transcript
    Why it matters
    Animation beats and cuts follow the spoken word, not a guess.
  • Review frames

    Value
    One frame every 7 seconds
    Why it matters
    A regular sample catches layout and brand problems across the whole film.

Subtitles are an option. video-use burns subtitles in 2-word uppercase chunks by default. If your named style has its own caption look, say so in the brief. HyperFrames has caption scenes in its catalog and a captions reference in /hyperframes.

Step 7

Hold the hard rules and run the self-check

The last line of Walid's prompt is a set of hard rules. Treat them as fixed. A brand film makes claims in your name, and an agent will fill gaps with plausible material if you let it.

The hard rules and why each exists
  • No fake screenshots

    Why it exists
    A mocked dashboard on screen reads as a real product or real results. If you have no real screenshot, draw the idea as a diagram in your named style.
  • No faces unless you provide them

    Why it exists
    The agent has no right to invent a person who looks like your team or your customer. Faces come from files you supply, with permission.
  • No revenue numbers unless you approve them

    Why it exists
    Revenue is the number most likely to be repeated, quoted and checked. Only figures on your approved proof list go on screen.
  • No claim on screen without a source

    Why it exists
    The film opens on a claim you fact-checked. Every other claim needs the same treatment, with the source visible.

Then comes step 7 of the prompt, the self-check. The agent must check its own work before it shows you anything. It transcribes the final mix and confirms every line is audible, extracts a frame every 7 seconds and reviews each one, then fixes what is broken. This works because the agent has tools to look and listen: video-use's transcript and filmstrip layers, plus your transcription service.

Prompt

Self-check pass

The film is assembled at {{final_path}}. Check it before you show it to me.

1. Transcribe the final mix. Compare it with the approved script line by line. List every line that is missing, cut off or hard to hear.
2. Extract one frame every 7 seconds. Review each frame. List anything wrong: overlapping text, text outside the safe area, wrong colors or fonts, a logo in the wrong state, a claim without a source line, a face or screenshot I did not provide.
3. Check the audio spec: cuts on word boundaries, 30ms fades at each cut, loudness at -14 LUFS. Report the measured loudness.
4. Write the issue list with timestamps.
5. Fix every issue, re-assemble, and run this check again.
6. Only when the list is empty, save the film to {{downloads_folder}} and open it. Show me the final issue list, even if it is empty.

Ask for the issue list every time, even when it is empty. An empty list you can see is better than a claim that everything is fine.

The last step of the prompt saves the result to your Downloads and opens it. Watch the film once with the sound on and once with the sound off. If the second view still makes sense, the visuals carry the idea on their own.

Watch out

Failure modes we see most

  • Unpinned tools change under you

    A render that worked yesterday breaks today because @latest moved. Fix: pin HyperFrames to a checked version, keep the version in the prompt and the project notes, and update the skills with npx hyperframes skills update to match.

  • A claim with no source

    The agent states a statistic it half remembers. Fix: fact-check the opening claim first, require a source line on screen for every claim, and cut any claim the agent cannot source.

  • Numbers you did not approve

    A revenue figure or client count appears in a scene. Fix: an approved proof list in the brief, and a stills review where you read every number.

  • Full render too early

    You wait on a render and then find the colors wrong. Fix: three stills per scene and no full render until you approve them.

  • Voice and animation drift apart

    Text lands before or after the word it describes. Fix: ask for word timestamps and named beat times, and sync every beat to them before assembly.

  • Lines you cannot hear

    Music or effects cover part of the voice. Fix: the self-check transcription. If a line in the script is missing from the transcript of the final mix, it was not audible. Re-mix and run the check again.

Time and money

Time and cost

Plan about three days: one for the script, one for the scenes and one for assembly and review. Scope decides the real number. A film with few scenes and a reference you already know goes faster. A film with several charts and a long fact-check goes slower.

Timeline

A planning estimate for the first film

  1. Script dayDay 1

    Inputs brief, fact-check of the opening claim, script and approval, three storyboards and your pick.

  2. Scenes dayDay 2

    Parallel scene builds, three stills per scene, your approvals and fixes.

  3. Assembly and review dayDay 3

    Voice with timestamps, sync, assembly, self-check pass, your final review.

A planning estimate, not a promise. Scope decides the real number.

Cost drivers
  • HyperFrames

    What it is
    Open source under Apache 2.0, free. Studio Pro is optional at $25 a month and adds video, image, voice and music models and 25,000 credits a month.
    Keep it down by
    Use the free CLI and bring your own voice and music keys, unless you want everything in one app.
  • Voice and music models

    What it is
    ElevenLabs, MiniMax, Lyria or ElevenLabs Music. Check current pricing on their sites.
    Keep it down by
    Generate the voice once from the approved script. Do not regenerate it for a scene fix.
  • Transcription

    What it is
    ElevenLabs Scribe or Deepgram. video-use uses Scribe for word-level timestamps. Check current pricing.
    Keep it down by
    Transcribe the final mix once per self-check pass.
  • Claude Code plan

    What it is
    Your existing plan. Sub-agents building scenes in parallel use more tokens than one agent.
    Keep it down by
    Approve the script and storyboard before the scene builds start. Use the Claude Code cost calculator to estimate.
  • Render compute

    What it is
    Local render on your machine, or hyperframes cloud render, or your own AWS Lambda.
    Keep it down by
    Render stills and previews while you iterate. Run the full render only after you approve the stills.

The biggest saving is process, not price. Each gate you hold keeps an expensive mistake from reaching a render. The agent can rebuild a scene in minutes. A review round that finds a wrong claim after the final render costs the whole assembly.

Results

How to measure it

Scorecard for an agent-made film
  • Approval rounds needed

    Where it comes from
    Count of rework requests at each of the four gates
    Which way you want it
    Down with each film
  • Time from braindump to final

    Where it comes from
    Timestamp of the brief vs. the saved MP4
    Which way you want it
    Down
  • Share of scenes approved at stills first time

    Where it comes from
    Scene log: approved at first stills vs. revised
    Which way you want it
    Up
  • Post engagement vs. your average

    Where it comes from
    Reactions, comments and reposts on the post, against your usual posts
    Which way you want it
    Up
  • Sourced claims on screen

    Where it comes from
    Count of claims with a source line, over all claims
    Which way you want it
    All of them
  • Issues found in the self-check

    Where it comes from
    Issue list from the first self-check pass
    Which way you want it
    Down as your prompt and brief improve

Take a baseline before you publish: your average reactions, comments and reposts on recent posts, and the time your last video took by whatever method you used. Re-measure at the first and second film. The approval-round and first-time stills numbers should improve as your brief and your named style get tighter. Walid's reference points are two posts: the brand film post with 34 reactions, 11 comments and 5 reposts, and the 3Blue1Brown explainer post with 50 reactions, 9 comments and 2 reposts, both as of 2026-11-22. Use your own numbers as the bar.

Questions

FAQ

What is HyperFrames?

HyperFrames is HeyGen's open-source AI video editor and framework, released under Apache 2.0. It turns HTML, CSS, JavaScript, media and seekable animations into deterministic MP4 videos. You can use it from the CLI, from coding agents such as Claude Code, Codex and Cursor with skills, or as a rendering core.

Can Claude Code make a video?

Yes, with the right tools. Claude Code writes the scenes as HyperFrames compositions, and video-use handles cuts, audio fades and word-level timing. You approve the script, a storyboard and still frames before any full render, and the agent checks the final mix itself.

How do I install HyperFrames for Claude Code?

Run npx skills add heygen-com/hyperframes and pick Core Skills, or add the plugin with claude plugin marketplace add heygen-com/hyperframes and claude plugin install hyperframes@hyperframes. You need Node.js 22 or newer and FFmpeg. Pin the CLI version instead of using @latest.

How do I make AI motion graphics that match my brand?

Give the agent your logo, colors and fonts as files, and name a visual style in one phrase, such as 3Blue1Brown with its traits. Add a reference film to match in structure and pacing. Review three still frames per scene before any full render.

How long does an AI explainer video take with this workflow?

Plan about three days: a script day, a scenes day and an assembly and review day. Scope decides the real number, and each approval gate you hold keeps expensive mistakes out of the final render.

Sources

AI Engineers

Want a repeatable film pipeline for your team?

An AY engineer joins your team, sets up the pinned toolchain, writes your inputs brief and prompt with you, and runs the first film through the approval gates. You keep the scenes, the prompts and the code.

From €5,000 a month per engineer. Starting price, your quote may be higher. First month money-back.