🔍 Read the full analysis: The Production Rules I Learned While Making An AI Video on ThorstenMeyerAI.com
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
TL;DR
Thorsten Meyer says his AI drawing engine produced a dull first full video, prompting him to measure frame use, stillness and camera movement and add checks to the build. The account is a personal report, not an independent evaluation; its figures come from his own notes and published experiments.
Thorsten Meyer has published an account of how he took an AI-generated whiteboard-video idea from an evening prototype to a revised production workflow, after the first full video felt slow and dull to him. The account describes a process involving independent AI code review, human approval and measurable checks; the evidence and results are based on Meyer’s own build notes and experiments, not an independent assessment.
Meyer says the idea began after he watched a hand-drawn explainer and wanted to create a similar broad visual approach without copying its drawings, assets or layouts. He says he wrote down a short list of general traits, including hand-drawn strokes, a continuous visual world and large numbers used as accents. Claude built a new engine, while OpenAI’s Codex reviewed the code. Meyer says Codex identified five issues in its first review, which were fixed alongside tests designed to fail on the earlier code.
A roughly one-minute test video won Meyer’s approval and was published in an article about local votes against data centers. The first full video produced the next morning did not meet his expectations: he described excessive white space, icons that did not support the narration and finished drawings that remained still. He says that video remains online unchanged as an example of what the team wanted to improve.
Rather than continue adjusting the video by feel, Meyer says the team measured frame coverage and the time images stayed completely still. On the first three scenes of the rebuilt video, he reports that filled-frame coverage rose from 15% to 66%, stillness fell from 29 seconds to zero, and camera moves increased from three to 27. He says the measurements became production gates, so scenes that are too empty can be stopped before rendering and objects that do not illustrate the narration can trigger a warning.
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 Video Critique to Production Checks
The account’s central point is that producing a video that renders is not the same as producing one that meets a defined standard. Meyer describes turning subjective feedback—too much empty space, too little movement and weak links between visuals and narration—into checks that can catch similar problems before publication. That approach gives readers a concrete example of how a small AI production workflow can combine automated review with human judgment.
The reported measurements show what changed on the first three scenes, but they do not establish that the revised video is better for every viewer or that the same workflow will work for other projects. The figures are Meyer’s own measurements, and the article provides no independent review or broader audience study. The practical takeaway is narrower: teams can define specific quality criteria, test whether their system meets them and keep a person responsible for approving the finished work.
Meyer also warns against treating a single metric as proof of quality. In a separate experiment described in the article, a model answered more tasks in a legal benchmark after a weight edit, but its legal score remained zero. That example reinforces the distinction between completion and correctness, a concern that applies to video production as well as model evaluation.
Products Worth Considering
Bose QuietComfort Headphones (2nd Gen), Wireless Over-Ear Noise-Cancelling Headphones, Black
Sony WH-1000XM6 Wireless Noise-Canceling Headphones, Black
LG UltraGear 27-inch QHD OLED Gaming Monitor, 240Hz
As an affiliate, we earn on qualifying purchases.
How the First Video Fell Short
Meyer frames the piece as a personal account of bringing AI ideas into production, including failures. He says the initial video style was inspired by another creator’s work, so the team set a clean-room boundary: it recorded general stylistic traits but used no drawings, assets or layouts from the original. The engine was built from scratch, according to Meyer, using his own design system and an openly licensed handwriting font.
The workflow separated creation from review. Claude built the engine; Codex, from another company, was assigned to inspect it rather than build it. Meyer says each genuine review finding led to a test that failed against the old code, which was intended to show that the test could detect the problem rather than merely pass after a fix. He also describes his own approval of the short test video as a required decision, not a substitute for automated checks.
The article’s updated note says its figures now refer to 10 published Lexassay tasks, with runs available on lexassay.com. Meyer’s related legal-model experiment is presented as another example of why output counts alone can be misleading: the model answered all 10 tasks, up from four, while its legal score remained zero.
“A demo that runs is not a decision.”
— Thorsten Meyer
Limits of the Reported Results
The article does not provide an independent evaluation of the video engine, the code review or the reported improvements. Meyer’s measurements cover the first three scenes of the rebuilt video, and the account does not specify a complete scoring method for frame coverage or camera movement. It also does not report viewer testing showing whether audiences found the revised video more engaging or clearer.
The source material describes checks intended to block empty scenes and flag visuals that do not match narration, but it does not give their false-positive rate or explain how often a human overrides them. The separate legal-model result likewise comes with a clear limitation: although the model answered all 10 published tasks, its legal score remained zero. The article does not establish why that score stayed at zero or whether the edited model performed better on other measures.
Further Runs and Human Review
Meyer says the updated Lexassay runs are public at lexassay.com, where readers can inspect the results for the 10 published tasks. The source does not identify a scheduled next release, a new benchmark milestone or an independent audit of the video workflow.
For now, the process described in the article relies on continued use of its production gates and a human’s approval of the finished video. Whether those checks prevent recurring problems across later projects, and whether viewers prefer the revised style, remains to be shown through further releases and evaluation.
Key Questions
What happened to the first full AI-generated video?
Meyer says it came out the morning after a short test video was approved. He found it slow and dull, citing excess white space, weakly connected icons and drawings that stopped moving. He says it remains online unchanged.
What changed in the rebuilt video?
Meyer reports that, across the first three scenes, frame coverage rose from 15% to 66%, stillness fell from 29 seconds to zero, and camera moves rose from three to 27. He says the measurements were turned into checks that can stop or flag scenes before rendering.
Who built and reviewed the video engine?
According to Meyer, Claude built the engine and OpenAI’s Codex reviewed the code. He says Codex identified five issues in the first review and that each fix came with a test that failed on the previous code.
Are the reported improvements independently verified?
No independent verification is described in the source. The figures come from Meyer’s own build notes and experiments, and the reported video measurements cover only the first three scenes.
What does the legal-model experiment show?
Meyer says a model answered all 10 published legal tasks after a weight edit, compared with four beforehand, but its legal score remained zero. He presents this as a warning that answering more tasks does not by itself prove correctness.
Source: ThorstenMeyerAI.com
