How a video gets made.
A person picks the topic. Claude researches it, writes the script and writes the code that draws every frame. Nothing is drawn by hand and nothing is stock footage.
One video, start to finish.
The production log of our next video, Everything Is Poison, shortened. Claude Opus reviews, Claude Sonnet builds, and the founder has the last word.
- Research
- Script
- Animation
- Voices
- Review and render
- Research
- Founder
Picks the topic: everything is a poison, so why is that not scary?
- Claude Opus 5.5
Grades 51 claims by strength of evidence, from A (solid) to D (do not say). Lists 13 things the video must never say.
- Script
- Claude Opus 5.5
Locks the script: 92 spoken lines, each with the picture it needs and a note on how to say it.
- Animation, in review rounds
- Claude Opus 5.5
Briefs the builder: a rough animatic first. To pass, timing must be clean, no voices may overlap and loudness must be even.
- Claude Sonnet 5.5
Writes the scene code and renders the animatic: 71 scenes, 7 minutes 55 seconds.
- Claude Opus 5.5
Round 3: rejected Five labels on the closing cards shrank below 44 pixels, too small to read on a phone.
- Claude Sonnet 5.5
Fixes it: all five labels back at 44 pixels, each inside its card.
- Claude Opus 5.5
Round 4: passed Loudness −14.00 LUFS, every spoken line within 0.4 dB of the others.
- Voices
- ElevenLabs
92 lines voiced, one take per line. Every scene is then re-timed to the real voice.
- Review and render
- Founder
Watches the cut and sends notes: the elbows look wrong when Zee drinks.
- Claude
Fixes the drinking pose: the elbow now hangs below the shoulder, like a real arm.
The same second of the video, rendered before and after the fix on 7 October. - Render
Final render: 8 minutes 37 seconds, −14.0 LUFS, English subtitles.
The five stages in detail
-
Research and script
- What happens
- Claude researches the topic and writes down each claim with its source and how sure we can be. Then it argues with its own draft from several seats: fact-checker, American viewer, animator, skeptical teacher. The script ends up as a table: each spoken sentence, the picture that goes with it, and a note on how it should sound.
- Tools
- Claude Code running Claude Opus. No code yet.
- What the human does
- Picks the topic, starts the session, reads the script.
-
Code the video, with a stand-in voice
- What happens
- Claude writes Python that draws every scene: characters, props, backgrounds, maps, charts, transitions. A shared engine and a library of drawings supply the parts, so Zee, the kitchen and the world map are reused, not redrawn. Timing comes from a stand-in voice, the text-to-speech built into macOS. Claude renders preview frames and looks at every scene. Two checkers then scan the layout: one finds covered text or text on a busy background, the other finds text drawn over pictures.
- Tools
- Claude Code running Claude Sonnet. The hardest scenes (maps, animated charts, many layers) get the same session switched to maximum effort. Python with Pillow and NumPy draws the frames, and ffmpeg assembles them.
- What the human does
- Starts the stage. On newer videos, also approves the rough cut with the stand-in voice before any paid voice is made.
-
Voices
- What happens
- Each sentence gets its own take, with delivery tags taken from the script’s tone notes. Hard names are spelled out phonetically. Then the scenes are re-timed to the real voice, because real takes run longer or shorter than the stand-in.
- Tools
- ElevenLabs text-to-speech, called from Claude Code. The narrator is Max and Zee has a separate voice. They are AI voices.
- What the human does
- Starts the stage. Claude tries three to five sentences first to settle the voice and the delivery, then does the rest.
-
Subtitles, titles and thumbnails
- What happens
- Subtitles are translated sentence by sentence into at least eight more languages and cut into short one-line cues. Claude writes at least five title options and builds at least five thumbnails, each from a different angle on the story, and checks that they still read at 320 pixels wide. Chapter marks in the description come from the timing file.
- Tools
- Claude Code: Sonnet for subtitles, Opus for thumbnails and translation. Thumbnails are drawn with the same character code as the video.
- What the human does
- Picks the title and the thumbnail.
-
Render and review
- What happens
- The final render only starts when the human says “render video”. Then a fresh Claude session, one that did not build the video, checks a frame about every three seconds and measures the loudness of every spoken line. Each line is balanced, then the whole track is normalized to −14 LUFS with a limiter.
- Tools
- Claude Code running Claude Opus, in a new session so the reviewer is not the builder. Renders go through a queue, one at a time, on a MacBook Air with 8 GB of RAM.
- What the human does
- Gives the render command, watches the finished video, reports small visual glitches, and uploads it to YouTube.
How a fact gets into a script.
Before anything is drawn, Claude Opus 5.5 researches the topic, grades every claim by how strong the evidence is, and argues with its own draft in a review round. The founder reads the result.
- A
- Solid: official or several independent sources.
- B
- Likely: a good source, numbers rounded.
- C
- Debated: the line has to say “may”.
- D
- Wrong or unproven: never said.
-
The source
“All things are poison, and nothing is without poison. Only the dose makes a thing not a poison.”
Paracelsus, 1538, in translation. Graded A.
-
The catch
The famous short version, “the dose makes the poison”, came later. He never wrote those exact words, so the video cannot say he did.
-
The line in the video
“Today we say it shorter: the dose makes the poison.”
One of 51 graded claims behind this video.
The do-not-say list matters as much. That video was never allowed to suggest that normal drinking is risky, so Zee says it out loud: No! Normal drinking is completely safe.
Watch a scene get built.
The classroom from our library, built by its own code, line by line. When it is finished, move your pointer over it: the layers move at different speeds, the way they do on camera.
the-graphite-mind/longform/engine/kitlib/scenes.py

One line, twelve languages.
A moment from our first video, with the subtitle that is on screen in each language. Claude translates every line, then cuts the text into short cues that are easy to read.
Uphill. Both ways.
Hundreds of drawings, none of them drawn by hand.
The shared drawing engine is 33,250 lines of Python in 83 files. 25,370 of those lines are the asset library: 314 props, 72 animals and plants, and 97 registered character designs, counting each outfit and age variant on its own.
Every video also has its own scene code on top of that, so these numbers are a floor.
Every long video and every Short on the channel is drawn by that engine.
Six frames from published videos.
All of it comes out of the engine. No stock clips.
What the human does
Picks the topics. Starts each stage. On newer videos, approves the rough cut with the stand-in voice before any paid voice is made. Picks the title and the thumbnail. Gives the render command. Reviews the finished video. Uploads it.
Claude Code does the typing, the drawing code and the checking. A person decides what gets made and what gets published.
What isn’t code
- The voices are AI voices made with ElevenLabs.
- World maps use Natural Earth data, which is public domain.
- The fonts inside the videos (Bradley Hand and Arial Rounded) are fonts already on a Mac.
- The music and sound effects are generated by code too.
The Graphite Mind is an independent channel. It uses Claude Code, a tool from Anthropic, and is not sponsored or endorsed by Anthropic.





