🔍 Read the full analysis: An AI Idea, One Boring First Video, And Lessons Learned The Hard Way 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 says he built a hand-drawn video engine in one evening, but its first full video was slow and visually sparse. He measured the problems and added checks to block scenes that are too empty or lack a clear link to the narration. The account is a personal report, not an independent evaluation of the engine.
Thorsten Meyer says he built a hand-drawn video engine in one evening, then found that its first full video was slow and visually boring the next morning. In a personal account published November 6 on ThorstenMeyerAI.com, he describes how he measured the problems and added automated checks and human approval before later videos went live.
Meyer says the idea came from a hand-drawn explainer video he liked. To avoid copying it, he says he wrote down a limited set of general style traits and built a separate engine: hand-drawn strokes, a continuous visual world, large numbers as accents, and colors from his own design system. He says no drawings, assets, layouts or drawing code were taken from the original. Claude built the engine, while OpenAI’s Codex reviewed the code rather than building it. Meyer reports that five findings from Codex’s first review were fixed that evening, each with a test designed to fail on the earlier code.
After Meyer approved a roughly one-minute test video, it appeared in an article on local votes against data centers. The first full video in the new style came the following morning. Meyer says it had too much white space, icons that did not support the narration, and finished drawings that remained still. He left that video online unchanged as an example of the shortcomings.
Instead of adjusting the result only by feel, Meyer says he measured frame coverage and time spent still. Across the first three scenes of a rebuilt video, he reports that filled-frame coverage rose from 15% to 66%, still time fell from 29 seconds to zero, and camera moves increased from 3 to 27. The account says these measures are now used as gates: checks can stop a scene that is too empty or flag an object that does not illustrate what is being said.
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 Gates
The account offers a practical example of a gap between a working AI-generated demo and a production-ready result. Meyer’s central point is that a video can render successfully while still failing as communication: its imagery may be sparse, static or disconnected from the narration. His response was to turn subjective criticism into measurable checks and to keep a person responsible for approving the output.
The approach also makes the limits of automated review visible. Meyer says Codex identified code issues, but the account does not present its review as a substitute for watching the video or judging its quality. A separate experiment he describes produced more completed answers on a legal benchmark without improving the model’s legal score. Together, the examples illustrate why a passing build or a higher completion count is not, by itself, proof that an AI system has improved in the way users need.
These are the author’s own results, drawn from his build notes and experiments. The account does not provide an independent audit or evidence that the same methods will produce similar results in other systems. Its value to readers is as a documented production process, including a failed first version and the checks added afterward.
Products Worth Considering
LG UltraGear 27GX704A-B 27-inch QHD OLED Gaming Monitor
Apple MacBook Pro 2021 14-inch with M1 Max Chip, 64 GB RAM, 2 TB SSD, Silver
Logitech MX Keys S Combo Wireless Keyboard and Mouse, Graphite
As an affiliate, we earn on qualifying purchases.
How the Video Engine Was Built
Meyer frames the video project as a production experiment rather than a general review of AI tools. He describes a division of labor: Claude generated the code, OpenAI’s Codex reviewed it, and Meyer made the human approval decision. He says every finding accepted from the first code review was paired with a test that failed on the old version, a way to check that the test could detect the reported problem.
The sequence matters to his account. A short test video received his approval, but the next morning’s full video exposed problems that the initial demo had not resolved. The later rebuild addressed those specific criticisms, and the resulting measurements became build rules rather than one-time observations. Meyer says an improved-look video about the Nasdaq’s worst day in 19 months went live a few hours after his critique; the source does not give a publication date for that video.
Meyer also connects the video experience to a separate experiment that week. He says a model completed 12 of 15 answers after a weight-editing experiment, compared with 4 of 15 before, while its legal score remained zero. That result is presented as another example of a metric that can look better without showing the intended improvement.
“And honestly, it was slow and boring.”
— Thorsten Meyer, describing the first full video
What the Report Does Not Establish
The article is a first-person account, and the performance figures come from Meyer’s own notes and experiments. It does not include an independent review of the video measurements, the engine’s code, or the legal benchmark. The source also does not specify the benchmark’s scoring method, what counted as a completed answer, or how the test set was constructed.
It is not clear whether the reported visual gates have been tested across a larger set of videos or how often they produce false alarms. The article gives results for the first three scenes of the rebuilt video, not a broader comparison across multiple projects. It also does not explain whether other reviewers approved the work or how the separate legal experiment was evaluated beyond the figures reported.
Further Tests and Published Results
Meyer says the video engine now uses frame coverage and narration-linked object checks as gates before rendering, and he points readers to an explainer video showing the process. The source does not give a schedule for further releases, independent testing or additional measurements, so those steps remain uncertain.
The legal-model experiment is described as a separate article, but the provided material ends before giving its full results or methods. Readers seeking to judge either project will need more detail on the benchmarks, the test coverage and whether the checks hold up across later videos and model runs. For now, the confirmed development is Meyer’s reported change in workflow: review the code, inspect the output, measure specific failures and use those measurements to block similar issues.
Key Questions
What happened to the first full video?
Meyer says the first full video in the new style was produced the morning after he built the engine. He found it too sparse and static, and left it online unchanged as an example of what did not work.
How did the rebuilt video change?
For the first three scenes, Meyer reports that filled-frame coverage increased from 15% to 66%, still time dropped from 29 seconds to zero, and camera moves increased from 3 to 27. These are figures from his own account, not an independent evaluation.
What roles did Claude and Codex have?
Meyer says Claude built the video engine and OpenAI’s Codex reviewed the code. He describes Codex as a reviewer, not a builder, and says accepted findings were paired with tests that failed on the old code.
Did the legal-model experiment improve the model?
Meyer says the model completed 12 of 15 legal benchmark answers after the edit, compared with 4 before, while its legal score remained zero. The source does not provide enough benchmark detail to independently assess that result.
Source: ThorstenMeyerAI.com
