You can make a finished explainer video by giving Claude Code one structured prompt and the HyperFrames skills. Our prompt below covers story, motion, sound, captions and checks. We ran it on 3 Nov 2026 and got a 29-second video with voice, music and captions, with no sign-in.
- What you need: Claude Code with the HyperFrames skills, FFmpeg, and an offline voice such as Kokoro.
- What you paste: the prompt below, with the topic, audience, outcome and length filled in.
- What came out of our test: a 28.8-second, 1920x1080 video with voice, quiet music and captions, made on one computer.
- What we had to fix: invisible captions, a voice that was too short, and music that needed a sign-in.
Think of a small film studio with one person in every job. The prompt names each job (director, writer, sound, editor), says what “finished” means, and tells the agent to check its own work before it hands the video over. You supply the idea; it does the rest.
What do you need before you start?
The plain point: three free-to-run tools and a coding agent, and no account for the voice or the music.
- Claude Code and the HyperFrames skills. HyperFrames is HeyGen’s open-source tool that turns a web page into an MP4, and its faceless-explainer workflow is the one for explaining a topic with no presenter on camera.1 Install steps, licence and the telemetry setting are on the HyperFrames page; read it first.
- FFmpeg and Node.js. The README lists “Node.js 22+, FFmpeg” as requirements.2 FFmpeg is a free program that joins pictures and sound into a video file.
- An offline voice. The explainer workflow can use Kokoro, a voice that runs on your own computer, instead of a signed-in online voice.3 Its model card calls it “an open-weight TTS model with 82 million parameters” and says “This is an Apache-licensed model.”4 TTS means text-to-speech; see text-to-speech.
- Room on the disk. The voice, transcription and background-removal tools may download a model the first time you use them.5 The prompt tells the agent to check free space first.
You also need a coding agent. Remotion’s documentation notes that “Most coding agents require a paid subscription.”6 This site lists no prices; check your agent’s own plan page.
What does each part of the prompt do?
The plain point: the prompt is thirteen labelled blocks. Only inputs is yours to fill in; the rest are rules the agent follows. The table says what each block is for.
| Part | What it does | Why it is there |
|---|---|---|
| <goal> | Does: says the deliverable is a rendered video with sound and captions, plus the files that re-render it. | Why: stops the agent handing back a plan or a single frame. |
| <role> | Does: makes the agent the whole studio, and tells it to log every choice it makes in DECISIONS.md. | Why: no stalling on small choices, and you get a record to review. |
| <principles> | Does: one idea per video, start from something real, cut what does not help, label made-up examples “Illustration”. | Why: keeps the video short and honest. |
| <constraints> | Does: no sign-in, no paid key, offline voice, check disk space, do not pause for approval. | Why: the run needs nothing from you and cannot spend money. |
| <inputs> | Does: holds your topic, audience, outcome, length and formats. | Why: the only block you must edit. |
| <discovery> | Does: the one question the video answers, what to show and what to cut, the opening hook. | Why: a video with no single question wanders. |
| <story> | Does: five beats with time shares: question, model, proof, turn, payoff. | Why: gives the script a shape before any design. |
| <visuals> | Does: colour, type and spacing first; neutral background, one accent; a list of banned effects. | Why: keeps the look plain and consistent. |
| <motion> | Does: movement must mean something; one thing moves at a time. | Why: viewers can follow one thing at once. |
| <sound> | Does: voice, script length, quiet music under the voice, at most three sound effects, never silent. | Why: an explainer without sound reads as unfinished. |
| <captions> | Does: burned-in captions from the final script, kept in the bottom band, plus a separate caption file. | Why: many people watch with the sound off. |
| <render_and_check> | Does: render the wide version, watch it end to end, check facts, captions and length, fix and re-render. | Why: a render is not the same as a correct video. |
| <deliver> | Does: lists the files to hand back and asks for a three-line summary. | Why: you can re-render the video later with one command. |
What is the full prompt?
The plain point: copy the whole block, fill in the brackets, and paste it into Claude Code in an empty folder. It is this site’s own wording. Its structure was inspired by a widely shared “Motion Studio” prompt; we rewrote it and added the sound, captions, render checks and constraints, and we replaced a line that told the model to imitate a named designer with plain principles of our own.Reasoning, 3 Nov 2026 — the prompt and this description of how it differs are this site’s own; no outside source is quoted.
How do you fill in the brackets?
The plain point: the brackets are all in the inputs block and the lines that repeat it. Write them the way you would brief a person.
- [TOPIC]: the question people ask, such as “What is a token in AI?”
- [AUDIENCE]: who they are and what they wrongly believe now. The prompt uses that wrong belief as the hook. We wrote “adults new to AI who think chatbots read whole words like people do”.
- [OUTCOME]: what they can do or explain afterwards, in one sentence.
- [SECONDS]: the length, 30 to 90. Shorter is easier to check.
- Formats: start with the 16:9 master. Add 9:16 and 1:1 only once the master passes.
- [slug]: a short file name, such as
what-is-a-token-in-ai.
Two edits are worth a thought. If you want to approve the video before it renders, delete the line “Do not pause for approval”: HyperFrames’ own command-line skill tells the agent to “Render only after approval”.7 Our prompt overrides that, because our test was meant to run on its own. And if you already have a voice account or brand files, remove the constraint that no longer applies.
We tested it: what happened on 3 Nov 2026
The plain point: it worked. The agent produced a 28.8-second video, “What is a token in AI?”, with an offline voice, a synthetic music pad and captions, rendered on our own computer with no sign-in and no paid tool.First-hand: video file metadata read with FFmpeg on 3 Nov 2026 — 1920x1080, 30 frames a second, H.264 video and AAC stereo audio, 28.77 seconds, 2,402,380 bytes. The test project’s own notes (SCRIPT.md, DECISIONS.md, SOURCES.md, README.md) record the rest of this section. The tool was HyperFrames 0.8.152. Press play, with captions on, to see it.
Read what the video says
You might think a chatbot reads whole words, like you do. It doesn’t. It reads tokens, small chunks of text. A short word like cat is one token. Long words are split. In English, a token is about three-quarters of a word. What you send, and what it writes back, share one limit: the context window. Paste a whole book to ask one question? That uses a lot. So when a chatbot stops halfway, it hit its token limit. Keep requests short.
What the agent did without asking
It followed the prompt’s story beats and wrote one line per decision in DECISIONS.md. It kept the look to the site’s own near-black, lime and white. It wrote a seven-line script of 82 words, made the voice with Kokoro (the “af_heart” voice), added three short sound effects, and burned in captions while also exporting a caption file. We told it to send no feedback report.
What we had to fix
- Caption contrast on a dark background. The first render’s captions were near-black text on a near-black background, a contrast of 1 to 1, so you could not read them. A style step had chosen the caption colour automatically and could not cope with a dark background. We set the captions to white. This is why the prompt now says to check caption contrast on dark backgrounds.
- Script length against the target. At normal voice speed the 82 words took only 25.2 seconds, under our 30 ± 3 second target. Slowing the voice to 0.88 speed, and rewording one line, gave 28.8 seconds. That is about 2.8 words a second, a little faster than the 2.5 the prompt suggests, so treat 2.5 as a starting guess and time the real voice.
- Music without a sign-in. The HyperFrames music library needs a HeyGen sign-in, and a local music model would have needed a download of more than 1 GB, which the prompt forbids. So the agent made a quiet 34-second synthetic pad with FFmpeg and set it low under the voice. It is plain, not memorable, but the video is not silent.
- Two smaller fixes. A check flagged an animation that moved spacing instead of position, and one line crossed the word chips on screen. Both were fixed and logged.
What we skipped: only the 16:9 version was made, not 9:16 or 1:1, and the step of showing two or three rough looks was skipped because the look was fixed in advance and the run was automatic. The facts in the video come from this site’s token page and the Google and OpenAI pages it cites; we did not re-read those pages for the test, so check any fact before you publish.First-hand: the test project’s SOURCES.md says its facts were copied from this site’s /meaning/token/ page (checked 1 Oct 2026), which cites Google and OpenAI, and were not re-fetched for the test.
Put the whole job, the rules and the checks into one prompt, then watch the finished video end to end yourself. The agent can make the video; checking that every fact and caption is right is still your job.
Questions people ask
Do I need to pay or sign in to use this prompt?
Not for the video tools or the voice: HyperFrames is Apache 2.0 software and Kokoro is an Apache-licensed model, and our test used no sign-in and no paid key. You still need access to a coding agent such as Claude Code, which usually has a plan of its own.8
How many words should the script have for a 30-second video?
About 75 to 85. The prompt suggests 2.5 spoken words a second, which is 75 for 30 seconds; our test used 82 words at a slowed voice and landed at 28.8 seconds. Always make the voice first, read its real length, then trim the words or adjust the pace.First-hand: 82 words, 25.2 seconds at normal Kokoro speed and 28.8 seconds at 0.88 speed, from the test project’s DECISIONS.md and the finished file, 3 Nov 2026.
Why does the prompt say not to wait for approval?
Because it was written for an unattended test run. For a real project you may prefer the opposite: delete that line, and the agent will stop at a preview, as HyperFrames’ own skill tells it to. Either way, watch the final video before you post it.
Where to go next
- “What is HyperFrames, and how do I install it?” See What Is HyperFrames?
- “What other video skills are there?” See AI video skills, sorted by job.
- “How do I brief any agent well?” See How to prompt AI agents, which has a ready-made brief for any multi-step job.
- “What do these words mean?” See token, text-to-speech and AI agent.
- “Is it safe to install a skill?” Use the pre-install checklist.
QUICK CHECK · 3 QUESTIONS
Pick an answer to see if you got it. Nothing is recorded.
1. Which part of the prompt is the only one you must fill in?
- The inputs block, which holds the topic, audience, outcome and length
- The constraints block, which lists the tools
- The captions block, which sets the colours
Show the answer
Answer: The inputs block, which holds the topic, audience, outcome and length. The page says only the inputs block is yours to fill in; the other blocks are rules the agent follows.
2. What was wrong with the captions in the first render of the test video?
- They were too small to read on a phone
- They were near-black text on a near-black background, so you could not read them
- They did not match the voice
Show the answer
Answer: They were near-black text on a near-black background, so you could not read them. The page says an automatic style step set the caption colour and could not cope with a dark background. The captions were changed to white.
3. How did the test get music without signing in to anything?
- It used the HyperFrames music library as a guest
- It downloaded a local music model of more than 1 GB
- It made a quiet synthetic music pad with FFmpeg and set it low under the voice
Show the answer
Answer: It made a quiet synthetic music pad with FFmpeg and set it low under the voice. The page says the music library needs a HeyGen sign-in and a local music model would be over 1 GB, which the prompt forbids, so the agent made a pad with FFmpeg.
- 1HeyGen, heygen-com/hyperframes README, “Creation workflows”, re-used from What Is HyperFrames?, read there 13 Oct 2026/faceless-explainer, “Explaining a topic / concept from arbitrary text”. ↑
- 2HeyGen, heygen-com/hyperframes README, re-used from What Is HyperFrames?, read there 13 Oct 2026“Requirements: Node.js 22+, FFmpeg”. ↑
- 3HeyGen, heygen-com/hyperframes skills/faceless-explainer/SKILL.md, re-used from What Is HyperFrames?, read there 13 Oct 2026voices resolve via “HeyGen (signed in)” or “Kokoro (offline)”. ↑
- 4Hugging Face, hexgrad/Kokoro-82M model card, read at source 3 Nov 2026“an open-weight TTS model with 82 million parameters” and “This is an Apache-licensed model.” ↑
- 5HeyGen, heygen-com/hyperframes skills/hyperframes-cli/references/upgrade-info-misc.md, re-used from What Is HyperFrames?, read there 13 Oct 2026tts, transcribe and remove-background, “Each may download its own model on first run.” ↑
- 6Remotion, Claude Code (docs), re-used from What Is HyperFrames?, read there 13 Oct 2026“Most coding agents require a paid subscription.” ↑
- 7HeyGen, heygen-com/hyperframes skills/hyperframes-cli/SKILL.md, re-used from What Is HyperFrames?, read there 13 Oct 2026“Render only after approval”; “Never render merely because checks pass.” ↑
- 8HeyGen, heygen-com/hyperframes LICENSE and README, re-used from What Is HyperFrames?, read there 13 Oct 2026“Apache 2.0 license, with no per-render fees or commercial-use thresholds.” ↑
Read at source 3 Nov 2026: Hugging Face, hexgrad/Kokoro-82M model card. Re-used with the date they were read on What Is HyperFrames? (13 Oct 2026): HeyGen, heygen-com/hyperframes (README, LICENSE, and the skills faceless-explainer and hyperframes-cli); Remotion, Claude Code (docs).
The test results are first-hand: this site ran the prompt on 3 Nov 2026 and read the finished file’s details with FFmpeg. The prompt, its table of parts, the “in plain terms” comparison and the advice marked as this site’s own are this site’s own, not claims from these sources.
Free printable AI cheat sheets, and AI news every week.
Sixteen one-page AI guides and checklists in one free PDF, no sign-up. And if you want to keep up, the newsletter covers what changed in AI each week, in plain English, with the sources.
SUBSCRIBE FREE ↗ GET THE CHEAT SHEET PACK