🔍 Read the full analysis: AI Project Takeaways From A Boring First Video on ThorstenMeyerAI.com
Get hardware and tech essentials delivered free with Prime
- Fast, free delivery on millions of items
- Prime Video, Amazon Music and more included
- Member-only deals all year
TL;DR
Thorsten Meyer describes how a hand-drawn video engine built with Claude produced a first full video he judged slow and boring. He says the team measured the problems, added checks to stop weak scenes, and used separate AI review and human approval; a related legal-model experiment also showed why completion counts alone can mislead.
Thorsten Meyer says a first full video made with a new hand-drawn animation engine looked slow and boring, prompting him to measure its visual problems and turn those measures into checks that can block later renders. In a personal account published on ThorstenMeyerAI.com and updated November 9, Meyer describes a workflow combining Claude for building, OpenAI’s Codex for code review, and his own approval before a video goes live.
Meyer says the idea came from an explainer video whose hand-drawn style he admired. To avoid copying it, he says the team wrote down a short list of broad style traits, including hand-drawn strokes and a continuous illustrated world. The account says the new engine used no drawings, assets, layouts or drawing code from the original. Its colors came from Meyer’s design system, and the handwriting font was openly licensed.
According to Meyer, Claude built the engine and OpenAI’s Codex reviewed the code without contributing to the build. Codex reported five findings in the first review round; Meyer says each was fixed that evening and paired with a test that failed against the old code. He presents that sequence as a way to check whether a test can detect the original problem rather than merely pass after a fix.
Meyer says he approved an approximately one-minute test video for publication that evening. The first full video, produced the following morning, drew his criticism for excessive white space, icons that did not support the narration and drawings that stopped moving when complete. He says that video remains online unchanged as an example of what the team wanted to improve.
Infographic 01 · The new video look
One evening, one boring first video, one measured fix
Speed was not the point. The steps in between were: a second AI that reviews, a human who says yes, and numbers that became gates.
Share of the frame filled with content on the first three scenes after the fix. Before: 15%.
From an evening idea to a measured fix
Relative times, this week
- EveningA video I likedClean room: only a list of style traits, no drawings, assets or layouts
- Same eveningOwn engine, first reviewClaude builds; OpenAI’s Codex reviews; 5 findings fixed, each with a test that fails on the old code
- Same eveningA human says yesOne-minute test video approved and published
- Next morningFirst full video: slow and boringToo much white space, icons that did not fit, a picture that stood still
- Hours laterMeasured fix, then gatesThe first video in the improved look goes live
Source: my own build notes for the new video look, this week (engine history, review round, approval, critique, before/after measurement of the first three scenes).
Powered by Thorsten Meyer AI
From Visual Critique to Build Checks
The account’s main point is that a working demo and a publishable result are different things. Meyer’s example describes a process in which human review follows automated checks: measurements can flag a scene, but a person still decides whether the finished video works. That distinction matters to teams using generative tools because output that renders successfully may still fail its intended purpose.
For the first three scenes of the rebuilt video, Meyer reports that the filled portion of the frame rose from 15% to 66%, time spent completely still fell from 29 seconds to zero, and camera moves increased from three to 27. He says these measurements became gates: the system stops a scene that is too empty and flags an object that does not illustrate what is being said. The figures apply to those three scenes, not necessarily to every scene or video.
Meyer also connects the production lesson to a separate experiment with a local model and legal tasks. He says it answered all 10 published benchmark tasks, up from four, while its legal score remained zero. The example underlines a limitation in treating task completion as a measure of quality: the count improved, but the score Meyer says mattered did not.
Products Worth Considering
Apple MacBook Air 13-inch Laptop with M5 Chip, 16GB Memory, 512GB SSD, Liquid Retina Display, and 12MP Center Stage Camera (2026, Silver)
LG UltraGear 34GX90SA-W 34-inch WQHD Curved OLED Gaming Monitor
LG UltraGear 25G590B-B 25-Inch Full HD 1000Hz Gaming Monitor
As an affiliate, we earn on qualifying purchases.
How the Video Workflow Developed
Meyer dates the initial idea and engine build to one evening, followed by a short test video and his approval. The next morning’s full video exposed problems that were less apparent in the demo. He says the team then measured frame coverage and stillness, revised the engine so camera movement followed the narration, and required objects to illustrate the spoken content.
The updated article says its figures refer to 10 published Lexassay tasks and that the runs are public on lexassay.com. Meyer says all figures in the piece come from his build notes, incident notes or published experiments. The article is a first-person account of his own development process, rather than an independent evaluation of the video engine or benchmark.
Meyer also describes an earlier monitoring failure from August: a guard generated about 1,300 warnings in a log that nobody read, while an alert email did not go out. He uses that incident to argue that monitoring has limited value if it does not reach a person. The account does not provide further details about the guard or the missed alert.
“The idea is the easy part.”
— Thorsten Meyer, in his November 6 article
Limits of the Reported Results
The reported video measurements cover only the first three scenes of the rebuilt video. Meyer does not provide a full-video comparison, an independent assessment of visual quality, or details about how the frame-coverage and stillness measures were calculated. The article also does not say how often the new gates block scenes or whether the checks have been tested across other projects.
The legal experiment’s reported figures do not include enough information in this account to assess the benchmark’s scoring method, the model, the weight-edit procedure or how the tasks were selected. Meyer says the runs are public, but the article’s summary alone does not establish that the model’s broader legal ability improved. The account also does not identify the one-minute test video’s viewers beyond Meyer’s own approval or report audience response to the published videos.
Further Runs and Review
Meyer says the improved animation approach was used for another video about the Nasdaq’s worst day in 19 months, published a few hours after his critique of the first full video. He says the video shown with his account uses the same method. The article does not give a schedule for further videos or say whether the engine’s quality checks will change.
The updated account points readers to public runs for the 10 Lexassay tasks. Further evidence about the approach would include results across more scenes and videos, clearer documentation of the measurement methods, and details about how the legal benchmark score is calculated. For now, the account presents the figures as Meyer’s own build and experiment results, alongside his description of the process that produced them.
Key Questions
What happened with the first AI-generated video?
Meyer says the first full video from his new hand-drawn engine looked slow and boring, with too much white space, weakly relevant icons and images that stopped moving after they were drawn. He says it remains online unchanged.
What changed in the rebuilt video?
Meyer says the team measured frame coverage and stillness, then added checks to stop scenes that were too empty and flag objects that did not match the narration. For the first three scenes, he reports coverage rose from 15% to 66%, stillness fell from 29 seconds to zero, and camera moves rose from three to 27.
Which AI tools did Meyer say he used?
Meyer says Claude built the engine and OpenAI’s Codex reviewed the code. He says Codex did not build the system, and that he personally watched and approved a short test video before publication.
Did the legal-model experiment prove the model became better at law?
No such conclusion is established by the reported task count alone. Meyer says the model answered 10 published tasks instead of four, while its legal score remained zero. The account does not provide enough benchmark detail to assess broader legal ability.
Source: ThorstenMeyerAI.com
