Source: raw/Claude_Makes_a_New_Kind_of_Video_-_Full_Tutorial.md

Creator: Futurepedia URL: https://www.youtube.com/watch?v=Ge1B-Ok1Pfk Duration: 11:19 Platform: YouTube Published: 2026-10-19 Skill level: operator

Futurepedia shows how to get motion-graphics explainer videos from an ordinary Claude chat, with no video model and no framework. Claude writes plain JavaScript that draws every frame, and the sound effects as well. The useful method is the style prompt that contains the full source code of an example video. Claude edits working code instead of guessing a style from a description or from screenshots. That is why, in his tests, Sonnet matched Opus. He built a free “style atlas” of 86 such styles, each with a 15-second sample, and walks through two runs: one from a topic only, one synced to a voiceover using a word-level SRT transcript. The atlas is free and the video has no sponsor segment.

This is the consumer-plan version of the “code” method in Claude Motion Graphics — three methods. It needs no Team or Enterprise plan, unlike Claude Motion, which the video doesn’t mention.

Key Takeaways

  • No framework. He used Remotion and HyperFrames before. “This doesn’t need any of that. Claude writes plain JavaScript that draws each frame by itself.” The output plays in the chat with its own sound effects, also generated in JavaScript. He credits a series of posts by another creator, Kevin, as the inspiration.
  • Put the code in the style prompt, not a video. “If you just send the video itself as a style reference, Claude and ChatGPT can’t watch it. All they do is take screenshots and analyze those.” His prompts are long (they paste as an attachment) because they include “all the actual code that was used to create the example video”, so Claude “just makes modifications instead of writing it all from scratch”. This matches the wiki’s note that Claude has no native video input without a skill such as watch.
  • Model choice matters less once the code is in the prompt. He ran the first example on Opus and says a Sonnet version looked “just as good”. He credits that to the embedded code. The same prompt in ChatGPT on GPT-6 Astra gave “a good result, too”, which he rates “a little worse than Claude’s”. The video doesn’t give model version numbers. These are one creator’s comparisons. On cost, see Sonnet 5.5 (about half Opus 5.5’s list price).
  • The style atlas. 86 styles: he listed 20 to 30 and had Claude invent the rest, along with a JavaScript sound-effect library. Examples include glassmorphism, 16-bit pixel, chalkboard, cut-paper diorama and flat vector science. Each has a 15-second “how AI works” explainer sample and a shorter style clip. The page has toggles for with voiceover (asks for a transcript), without voiceover (asks for a topic), and storyboard first. Copy Prompt combines the style code with your inputs.
  • Storyboard first is optional. With it on, Claude returns still images of each major beat before the full build, so you can change direction cheaply. He usually turns it off and asks for edits afterwards, because the sample already shows the style.
  • It uses a lot of tokens. Building the atlas was “the fastest I have ever burned through my usage and resets”. Generating a whole style library is expensive. Running one style prompt is a normal chat turn.
  • Edits are precise because it is code. “If you want any changes… it’s really easy for it to fix or alter anything really precisely.” Change the palette or the pacing (“things to move slower”) in the prompt or in follow-up turns.
  • Turn the style you like into a skill. Once you have a custom version you like, “ask Claude to turn that into a skill so you don’t have to come back to this database every time.” For the refinement loop behind a production skill, see animate skill playbook.

Voiceover-synced workflow

  1. Make the voiceover. He normally uses his own voice. For the demo he used Gemini 3.8 Flash TTS in Google AI Studio (model picker → Audio → Gemini 3.8 Flash TTS → Create new dialogue), chose a voice, pasted the script, set the style to Narration and ran it. Optional tags add breaths and laughs, which he says suit dialogue more than narration. ElevenLabs or any other method works. See the September 2026 model landscape for Gemini 3.8.
  2. Get a timestamped transcript. Some tools (ElevenLabs) return one with the audio. Otherwise use a free online or local transcription tool, or drop the audio into Claude: “Give me a word level SRT transcript for this that I can copy paste.”
  3. Word-level vs line-level. Word-level timing syncs best, but for longer videos he asks for a plain SRT (timed by line), because word-level files get “gigantic”. He uses word-level for short pieces.
  4. Build the prompt. In the atlas, pick a style (he used flat vector science for “how stars are born”), toggle With voiceover, type the topic, paste the SRT, optionally turn off the storyboard, then Copy Prompt.
  5. Paste into a new Claude chat, then attach the audio. The finished output has a Load voiceover control on the right: pick the file or drag it in. You can attach the audio with the prompt instead. He made sure it works afterwards because he forgot the audio on his first test.
  6. Refine the look over time. For a channel, spend time on one house style. His own is based on his studio background of blue and amber dots and lighting.

Try It

  1. Open the free style atlas (first link in the video description), pick a simple style such as glassmorphism, choose Without voiceover, enter a topic, and paste the copied prompt into a new Claude chat on Sonnet.
  2. Run the same prompt on Opus and compare the two outputs and the usage each one took on your plan.
  3. For a narrated piece under a minute: record or generate the voiceover, ask Claude for a word-level SRT, run the With-voiceover prompt, then load the audio in the output.
  4. When a variant fits your brand, ask Claude to turn it into a skill (style code plus your palette and pacing rules), then reuse it with / instead of the atlas.
  5. To make your own reusable style without the atlas, keep the code of a video you like and paste it as the reference next time, instead of a screen recording.

Implementation

  • Tool/Service: Claude chat (Opus or Sonnet). It also worked in ChatGPT. Futurepedia’s free style atlas is a web page that builds the prompts. Voiceover from Gemini 3.8 Flash TTS (Google AI Studio), ElevenLabs or your own recording.
  • Setup: None beyond a Claude account. The output plays in the chat, and the voiceover loads into it.
  • Cost: Not stated. Prompts are long because they carry full source code, so each run uses more input tokens than a short prompt would.
  • Integration notes: The video doesn’t show how to export to MP4. That matters if you are editing the result into a timeline (see Open Questions). Claude Motion (Team and Enterprise) exports MP4 natively. HyperFrames and Remotion render locally.

Open Questions

  • Export to a video file is not shown. The output plays in the chat with a voiceover loader. The video doesn’t say whether he screen-records it or renders it to MP4 some other way.
  • Which Claude surface renders it. It is presumably an artifact in the chat; the transcript says only “the option on the right”.
  • Model versions are unnamed. Uploaded 2026-10-19, so likely Opus 5.5 and Sonnet 5.5, but the video doesn’t confirm it.
  • The atlas URL is only “the first link in the description”. It wasn’t fetched, so the 86-style count and its licence aren’t verified.
  • The creator “Kevin” whose posts inspired the method isn’t identified further.