🔍 Read the full analysis: AI Video Development: The Rules I Learned Through Trial And Error on ThorstenMeyerAI.com
Get audio and creator gear 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 AI video engine in one evening, but its first full video was slow and visually sparse. He describes revising it using measurements, independent AI code review, human approval and automated gates, while cautioning that completion metrics do not prove quality.
Thorsten Meyer has described how he built and revised a hand-drawn AI video engine, turning an initial style experiment into a workflow with independent code review, human approval and measurable quality checks. In a personal account published November 6 and updated November 9, Meyer says the first full video was slow and visually dull; measurements of its scenes informed changes that later became automated build gates.
Meyer says the idea came after watching a hand-drawn explainer that he liked. To avoid copying it, he says he set a clean-room rule: record only broad style traits, such as hand-drawn lines and a continuous visual world, and use no drawings, assets, layouts or drawing code from the original. He says the engine was built from scratch, using his own design system and an openly licensed handwriting font.
In his workflow, Claude builds and OpenAI’s Codex reviews the code without building it. Meyer says Codex identified five issues in its first review, which he fixed that evening. Each fix, he says, came with a test designed to fail against the earlier code. He also says his own approval is required before a video goes live: after watching a roughly one-minute test video, he approved it for publication inside an article about local votes against data centers.
The first full video in the style, produced the next morning, did not meet his expectations. Meyer says it had too much white space, icons that did not support the narration and images that stopped moving after they were drawn. He measured the first three scenes and reports that the proportion of the frame filled with content rose from 15% to 66% after a rebuild; stillness fell from 29 seconds to zero, while camera moves rose from three to 27. He says those measures now act as gates that can stop sparse scenes or flag objects that do not illustrate the narration.
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 Quality Became a Build Check
Meyer’s account illustrates a practical distinction for teams using generative tools: producing a file is not the same as producing work that is useful or ready to publish. His described process combines automated checks with human judgment, rather than treating a successful render or a positive first reaction as sufficient evidence of quality.
The reported scene measurements give his team specific criteria to monitor, but they do not establish that every viewer will find the revised video engaging or that the same thresholds will suit other projects. The broader relevance is the method he describes: turn specific criticism into observable measures, then make recurring failures detectable before publication. The account also highlights that review, testing and approval each serve different purposes; none alone guarantees a strong result.
Products Worth Considering
Sony WH-1000XM5 Premium Noise Cancelling Wireless Headphones – Midnight Blue
Jackery Explorer 300 Portable Power Station with 40W Solar Panel
Sonos Arc Ultra Soundbar with Dolby Atmos, Black
As an affiliate, we earn on qualifying purchases.
From Style Test to Published Video
Meyer frames the work as a personal account of taking AI ideas into production, not as an independent evaluation of the tools. He says the first short test was published within an article about local opposition to data centers. The following morning’s full-length attempt was left online unchanged as an example of what he considered unsuccessful; a later, rebuilt video about the Nasdaq’s worst day in 19 months used the revised approach.
The account also describes a separate experiment with a local model and a legal benchmark. Meyer says a weight-editing experiment increased the model’s answers on 10 published tasks from four to all 10, while its legal score remained zero. He uses this result to argue that a higher completion count can be misleading when the measure that represents task quality does not improve. The updated article says its figures refer to the 10 published Lexassay tasks and that the runs are public on lexassay.com.
“A demo that runs is not a decision.”
— Thorsten Meyer
What the Report Does Not Establish
The reported figures come from Meyer’s own build notes, incident notes and experiments, according to the article. The account does not provide an independent evaluation of the video metrics, the legal benchmark or the overall quality of the finished videos. It also does not specify how the frame-fill and stillness thresholds are set, how they perform across different types of scenes, or whether viewers found the revised videos more effective.
The article describes a review setup involving Claude and OpenAI’s Codex, but does not establish that either system will identify every defect. Meyer also recounts an earlier monitoring failure: in August, he says, a guard wrote about 1,300 warnings to a log that nobody read, and an alert email did not go out. The account does not detail the later monitoring changes or show how consistently they reach a person.
Checks Before the Next Render
Meyer says the scene measurements have become gates in the video build, so the process can stop a scene that is too empty and flag an object that does not match what is being said. The workflow he describes also retains code review, tests that expose earlier defects, and a human approval step before publication.
The article does not announce a product launch or a schedule for further releases. Its next practical test is whether the checks continue to catch the problems they were designed for in future videos—and whether the resulting work holds up to human review. Meyer’s account presents that as an ongoing production process, not proof that the workflow has solved video quality across projects.
Key Questions
What did Meyer report about his AI video project?
He described building a hand-drawn video engine, finding its first full video dull, and revising the workflow using scene measurements, code review and human approval.
How did the video measurements change?
For the first three scenes of the rebuilt video, Meyer reports that frame fill rose from 15% to 66%, stillness dropped from 29 seconds to zero, and camera moves increased from three to 27.
Who built and reviewed the engine?
Meyer says Claude built the engine and OpenAI’s Codex reviewed its code. He says he fixed five findings from the first review and added tests for those issues.
Does the account independently verify the results?
No independent verification is described. Meyer says the figures come from his own notes and experiments; the account does not establish how well the metrics predict audience response or apply to other video projects.
What is the main caution from the legal-model experiment?
Meyer says the edited model answered all 10 published legal tasks, up from four, while its legal score remained zero. The result shows why a higher completion count alone does not establish better performance.
Source: ThorstenMeyerAI.com
