🔍 Read the full analysis: What One Evening Of Building An AI Video Taught Me 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, but its first full video was slow and visually dull. He revised the system using measured scene checks, added gates to catch problems before rendering, and says a human still has to approve the result.
Thorsten Meyer says he built a hand-drawn AI video engine in one evening, then revised it after its first full video proved slow and visually dull. In a November 6 account updated November 9, he describes how a second AI reviewer, human approval and measurable production checks shaped the system before a later video went live.
Meyer says the project began with a video style he admired: a hand-drawn explainer moving through a continuous whiteboard-like world. To avoid copying the original work, he says he wrote down a limited set of style traits and built from scratch, using no borrowed drawings, assets, layouts or drawing code. The engine used his own design system, including blue imagery, orange accents and an openly licensed handwriting font.
In his workflow, Claude built the engine, while OpenAI’s Codex reviewed the code but did not build it. Meyer says Codex identified five issues in its first review; he fixed them that evening and wrote a test for each finding that failed against the earlier code. He also says he watched a roughly one-minute test video and personally approved it for publication that evening, within an article about local votes against data centers.
The next morning’s first full video did not meet his expectations. Meyer describes it as having too much empty space, icons that did not support the narration and finished drawings that remained still. He says the video remains online unchanged as an example of what went wrong. After measuring the opening three scenes, he reports that the rebuilt version raised the filled-frame share from 15% to 66%, reduced completely still time from 29 seconds to zero and increased camera moves from three to 27. Those figures apply to the first three scenes, not necessarily the entire video.
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
Turning Video Critique Into Checks
The account illustrates a practical distinction for people using generative AI in production: a system that can produce a working demo may still produce work that is not ready to publish. Meyer’s first full video exposed problems that a successful build alone did not catch, including pacing, composition and whether visuals matched what was being said.
His response was to make parts of that critique measurable and use them as pre-render gates. According to Meyer, the engine now stops a scene if it is too empty and flags an object that does not illustrate the narration at that moment. The reported figures are from his own project notes, rather than an independent evaluation. Still, they show how a subjective reaction can be converted into project-specific checks, while leaving the final judgment to a person.
Products Worth Considering
Amazon Kindle Paperwhite Signature Edition (12th Generation, 2024), 32 GB, Metallic Raspberry
Samsung Galaxy Tab S7 11-inch Android Tablet 128GB Wi‑Fi with S Pen Mystic Black
Apple AirPods Pro 3 Wireless Earbuds with Active Noise Cancellation
As an affiliate, we earn on qualifying purchases.
From Style Reference to Published Video
Meyer frames the account as a personal report on taking AI ideas into production, rather than a controlled study. He says the first approved test video went live the same evening the engine was built. The first full video followed the next morning; after Meyer recorded his criticisms and the engine was revised, an improved video about the Nasdaq’s worst day in 19 months was published a few hours later. The account does not give a calendar date for those steps beyond the November 6 publication date.
The same article describes a separate experiment that Meyer says raised a related measurement concern. After Claude edited a local model’s weights for a legal task, the model answered all 10 published legal benchmark tasks, up from four, but its legal score remained at zero. The article’s November 9 update says the runs are public on Lexassay.com. Meyer presents this as an example of why completion counts alone may not show whether a system performs the task correctly.
“A demo that runs is not a decision.”
— Thorsten Meyer
Limits of the Reported Measurements
The figures come from Meyer’s own build and incident notes, as he states in the source account. No independent assessment, comparison with other video tools or evaluation of viewer response is provided. The frame-fill and motion numbers cover the first three scenes of the rebuilt video; they should not be read as measurements for every scene or as proof that audiences preferred the revised version.
The account also does not specify the exact date of the build, the full technical method used to calculate its metrics, or whether the new gates have caught errors in later production runs. The legal benchmark result is a separate experiment, and the article does not establish that its findings apply to other models or tasks.
Further Runs Will Test the Gates
Meyer says the video engine now uses the measured checks to stop scenes that are too empty and flag visuals that do not match the narration. The source does not set out a schedule for future videos or say whether the checks will be changed as more examples are produced.
For the separate legal-model experiment, Meyer says the runs for the 10 published tasks are public on Lexassay.com. Further results would help show whether the model can improve its legal score, not only complete more tasks. For now, the article’s next test is whether the production checks continue to identify problems while a human reviewer decides whether the finished video is fit to publish.
Key Questions
What happened in Meyer’s AI video project?
Meyer says he built a hand-drawn video engine in one evening, found its first full video slow and visually dull, then revised the system and added measurable checks before rendering.
What changed in the rebuilt video’s first three scenes?
According to Meyer, filled-frame share rose from 15% to 66%, completely still time fell from 29 seconds to zero, and camera moves increased from three to 27. These figures cover the first three scenes only.
What roles did Claude and Codex have?
Meyer says Claude built the engine and OpenAI’s Codex reviewed the code. He says Codex did not build the system; five findings from its first review were fixed, each with a test that failed on the previous code.
Did the video checks prove the finished work was good?
No. The reported checks measure selected features, such as how much of the frame is filled and whether the image is still. Meyer says a human must still review and approve the result, and the source provides no independent quality assessment.
What was the separate legal-model result?
Meyer reports that after a weight-edit experiment, a model answered all 10 published legal tasks instead of four, while its legal score remained at zero. The account cautions that completing tasks did not mean the answers were legally correct.
Source: ThorstenMeyerAI.com
