🔍 Read the full analysis: Taking AI From Concept To Production, One Evening At A Time on ThorstenMeyerAI.com
Get business pricing on tech for your team
- Business-only prices and quantity discounts
- Tax-exempt purchasing
- Multiple users, one account, clear invoices
TL;DR
Thorsten Meyer says he built a hand-drawn video engine in one evening, then revised it after the first full video proved slow and visually static. His account describes code review by a second AI, human approval, and production checks based on measured video qualities; the figures are from his own build notes and experiments.
Thorsten Meyer says he turned an idea for a hand-drawn AI video style into a working drawing engine in one evening, then changed the system after its first full video came out slow and visually static. In an account published on ThorstenMeyerAI.com and updated November 9, he describes using a separate AI code reviewer, human sign-off and measurable checks to decide whether the output was ready to publish.
Meyer says the initial concept came from a video style he admired, but he set a clean-room rule to avoid copying its drawings, assets, layouts or code. He says the team recorded a short list of general traits, including hand-drawn strokes and a continuous visual world, and built a new engine from scratch. The colors came from his own design system, and the handwriting font was openly licensed.
In his setup, Claude built the engine and OpenAI’s Codex reviewed the code without making changes. Meyer says Codex identified five issues in the first review, which were fixed that evening. He says each finding was paired with a test that failed against the old code, an approach intended to show that the test could detect the original problem rather than merely pass after a fix.
Meyer says he watched a test video of about one minute and approved it for publication that evening. The next morning, however, the first full video in the new style struck him as slow and boring. He recorded concerns about empty space, icons that did not support the narration and drawings that stopped moving once completed. That video remains online unchanged as an example of an unsatisfactory result, according to his account.
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
How Video Checks Became Production Gates
The account focuses on a practical risk in deploying generative tools: an output can render successfully without meeting the intended standard. Meyer says he responded by measuring frame coverage, stillness and camera movement rather than continuing to adjust the video by taste alone.
For the first three scenes of the rebuilt video, he 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 those measures were then turned into checks: scenes judged too empty are stopped before rendering, and the build flags objects that do not illustrate the spoken content.
Those figures describe a small set of scenes in Meyer’s own project, not an independent evaluation or a general benchmark for AI video systems. Still, the account offers a concrete example of combining automated review with human judgment and tests tied to observed problems. Meyer argues that the person responsible for a result must still watch it and approve it; a successful demo alone is not a release decision.
Products Worth Considering
Jackery Explorer 2000 v2 Portable Power Station with Two 200W Solar Panels
LG UltraGear 34G600A-B 34-Inch Curved WQHD Gaming Monitor, 160Hz
Amazon Kindle Paperwhite (12th Generation, 2024), 16GB, 7-Inch Display, Raspberry
As an affiliate, we earn on qualifying purchases.
The First Video Exposed the Gap
Meyer dates the original idea and engine build to an evening after watching a hand-drawn explainer. He says the roughly one-minute test version was published that same evening inside an article about local votes against data centers. The first full video followed the next morning, and its shortcomings prompted the redesign.
He also connects the production lesson to a separate experiment described on his site. In that test, Claude edited the weights of a local model for legal tasks. Meyer reports that the model answered all 10 published Lexassay tasks, up from four, but says its legal score remained zero. The figures refer to the 10 published tasks, and the runs are public on lexassay.com, according to the site’s November 9 update. The result illustrates why a higher task-completion count does not, on its own, establish that a system performs correctly.
Meyer’s account also mentions an earlier monitoring failure: in 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 is useful only if warnings reach someone able to act. These examples come from his own notes and experiments, not an external audit.
“A demo that runs is not a decision.”
— Thorsten Meyer
What the Report Does Not Establish
The reported video measurements cover the first three scenes of the rebuilt production, and Meyer does not provide an independent assessment, a larger comparison set or a full account of how the measurements were calculated. The article also does not establish whether the same gates work equally well across other topics, video formats or production teams.
The legal-model result is likewise limited to the published benchmark tasks described by Meyer. He says the model’s legal score remained zero despite answering all 10 tasks, but the account does not establish broader legal capability or explain the full scoring methodology here. The reason the August alert email failed is also not detailed in the supplied account. These are the author’s reported project results and interpretations, rather than independently verified findings.
More Runs and Checks Ahead
Meyer says the Lexassay runs are public on lexassay.com and that the updated figures refer to all 10 published tasks. The article describes the video checks as part of the engine’s production process, but does not announce a release schedule, product launch or independent validation effort.
The next useful evidence would be further published runs and results showing how the video gates perform across more scenes and projects, along with details on how alerts reach a responsible person. Until then, readers can treat the measurements as a documented account of one creator’s workflow, not as proof that the approach will produce the same results elsewhere.
Key Questions
What did Thorsten Meyer build?
He says he built a hand-drawn video drawing engine from scratch after developing an idea for a new visual style. The first working version was made in one evening, according to his account.
How did Meyer revise the first full video?
After describing the video as slow and visually static, Meyer says he measured frame coverage, stillness and camera movement. He reports using the resulting checks to stop scenes that are too empty and flag visuals that do not match the narration.
What role did the second AI play?
Meyer says Claude built the engine, while OpenAI’s Codex reviewed the code and reported problems. Codex did not build or modify the system in his workflow.
Do the reported results prove the system is reliable?
No. The video figures cover the first three scenes in one project, and the legal experiment involved 10 published tasks. Meyer reports that the legal score stayed at zero, and the account does not provide independent validation of either result.
Source: ThorstenMeyerAI.com
