TL;DR
Get the latest gadgets delivered free with Prime
- Fast, free delivery on millions of items
- Prime Video, Amazon Music and more included
- Member-only deals all year
Thorsten Meyer says he built a hand-drawn video engine in an evening, then improved its first full video after measuring scene composition and motion. His account describes using one AI to build, another to review, human approval before publishing and automated checks to catch repeat problems.
Thorsten Meyer says he built a new hand-drawn video engine in an evening, but its first full video was slow and visually sparse, prompting him to measure the problem and add checks before publishing. In an account published November 6, Meyer describes a workflow that pairs AI-generated code with a separate AI review, human approval and production gates.
Meyer writes that the idea came from a video style he admired, but he set a clean-room rule: record only broad traits such as hand-drawn strokes, a continuous illustrated world and large color accents. He says the team used no drawings, assets, layouts or drawing code from the original. The engine was built with Claude, using his own design system and an openly licensed handwriting font.
For the first code review, Meyer says OpenAI’s Codex examined the work without building it. Five findings were fixed that evening, each paired with a test that Meyer says failed against the old code. He then watched a roughly one-minute test video and approved it for publication. The following morning, the first full-length video looked too empty, used icons that did not support the narration and left the image still after drawings finished.
After revising it, Meyer reports that across the first three scenes, the share of the frame filled with content rose from 15% to 66%, time spent standing still fell from 29 seconds to zero, and camera moves rose from three to 27. He says the team turned two checks—whether a scene is too empty and whether an icon illustrates nothing being said—into gates that stop a build. These are figures from his own build notes, not an independent evaluation.
From prototype to production rules learned through review
Thorsten Meyer describes building a hand-drawn video engine in an evening, then improving its first full video by measuring scene composition and motion. His workflow adds separate AI review, human approval and build checks.
From idea to first release
Meyer says the process moved from a style description to a short test, code review and publication.
Record broad traits, such as hand-drawn strokes and large color accents, without reusing the original’s drawings, assets or layouts.
Claude built the engine; OpenAI’s Codex reviewed the code and reported five findings that Meyer says were fixed with tests.
Meyer watched a roughly one-minute test video and approved it for publication. The full video revealed issues the short test had missed.
Criticism became build checks
The first full video showed problems in three areas; Meyer says checks for two now stop a build when triggered.
- Scenes had too much empty space.
- Icons did not support the narration.
- The image stayed still after drawings finished.
- Flag scenes that are too empty.
- Flag icons that illustrate nothing being said.
- Use the checks as gates that stop a build.
What the figures can show
The reported before-and-after measures apply only to the first three scenes.
- The figures come from Meyer’s own build notes.
- The account does not give the measurement method or gate thresholds.
- It reports no independent rating or viewer-response data.
- Human review and approval remain part of the described workflow.
From Video Critique to Build Gates
The account shows how a quick AI-assisted prototype can fall short when judged as a finished video. Meyer’s process converts specific criticism—empty space, weak icon choices and static scenes—into measurable criteria and checks intended to catch repeat issues before publication.
His examples also make clear that review and monitoring depend on people and working systems. Meyer says a separate AI reviewer found issues despite the author’s tests passing in another experiment, and describes an earlier guard that produced about 1,300 warnings in a log nobody read while its alert email failed to send. In his account, green tests and recorded warnings did not by themselves establish that a result was sound or that anyone would respond.
Products Worth Considering
Sonos Era 100 SL Compact Microphone-Free Speaker, White
Bose SoundLink Flex Bluetooth Speaker (2nd Gen), Black
LG UltraGear 32GS60QC-B 32-inch Curved QHD Gaming Monitor (180Hz, 1ms, 1000R, AMD FreeSync, HDR10)
As an affiliate, we earn on qualifying purchases.
How the First Video Changed
The workflow began with an evening idea and a short style description, followed by an engine, a code review and a brief test video. Meyer says he approved that test and published it within an article about local votes against data centers. The next morning’s full video exposed problems that the short sample had not revealed.
He then assessed the first three scenes and recorded three measures: frame occupancy, seconds without movement and camera-move count. The reported before-and-after values apply to those scenes. Meyer says the resulting thresholds now stop builds when scenes are too empty or icons do not illustrate the narration. The published account does not provide the full measurement method or the exact threshold values.
“The idea is the easy part.”
— Thorsten Meyer
Limits of the Reported Measures
The measurements come from Meyer’s own notes and cover only the first three scenes. The account does not specify how frame occupancy or camera moves were defined, disclose the gates’ numerical thresholds, or compare the revised video with an independent rating. It also does not report whether the revised format performed better with viewers.
The workflow examples are personal accounts and experiments, not independent evidence that using two AI systems or these particular gates will improve other production pipelines. The article says the video was live, but gives no audience data or details about how the system will be maintained as the engine changes.
How the New Checks Will Be Used
Meyer says the checks for overly empty scenes and unsupported icons are now part of the build process and will stop production when triggered. That makes future builds subject to the rules he established after the first full-video critique.
His account does not name a next release date or describe a planned independent review. The next evidence will come from how the checks perform on later videos and whether human reviewers continue to approve the output before publication.
Key Questions
What did Thorsten Meyer build?
He says he built a hand-drawn video engine in an evening using Claude, based on a written description of style traits rather than copied drawings, assets or layouts.What was wrong with the first full video?
Meyer says it had too much white space, icons that did not support the narration and long stretches with no camera movement.What changed in the revised version?
For the first three scenes, his notes report that frame content rose from 15% to 66%, still time fell from 29 seconds to zero, and camera moves increased from three to 27.Did an AI reviewer approve the code?
Meyer says Codex reviewed code built with Claude and reported five findings that were fixed with tests. He separately says he watched and approved a roughly one-minute test video before it was published.Are the measurements independently verified?
The article attributes the figures to Meyer’s own build notes. It does not describe an independent verification or provide viewer-response data.Source: Thorsten Meyer AI
