🔍 Read the full analysis: What Building A Simple AI Video Taught Me About Production 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 describes building a hand-drawn AI video engine and improving its first full-length output after finding it slow, sparse and visually static. He says the changes were guided by measurements and turned into automated checks, alongside review by a second AI and human approval.
Thorsten Meyer says he built a hand-drawn AI video engine in one evening, then revised it after its first full video proved slow, sparse and visually static. In an account published November 6 on ThorstenMeyerAI.com, he describes how a second AI reviewer, his own approval and production checks based on measured results helped move the format from a quick demo to a repeatable workflow.
Meyer says the initial idea came from watching a whiteboard-style explainer whose look he admired. To avoid copying it, he says he wrote down a limited set of general traits—such as hand-drawn strokes and a continuous visual world—and used no drawings, assets or layouts from the original. Claude built the engine, according to Meyer; he says the palette came from his own design system and the handwriting font was openly licensed.
He describes a separate role for OpenAI’s Codex: reviewing the code rather than building the engine. Codex identified five issues in its first review, Meyer says, and each fix received a test intended to fail against the earlier code. He also says he watched a roughly one-minute test video and approved it before it appeared in an article about local votes against data centers. Meyer presents that approval as a human decision point, not an automatic consequence of a successful render.
The first full video in the new 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 once their drawings were complete. He measured the frame area occupied by content and the time the image remained still, then changed the engine. On the first three scenes of the revised 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. He says those measures became checks that can stop an overly empty scene or flag an object that does not illustrate the words being spoken.
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 Build Checks
The account offers a practical example of a broader production question: how to tell whether an AI-generated output is ready for viewers, rather than merely able to run. Meyer’s description separates generation, review and approval into distinct steps. That matters to teams using generative tools because a technically successful render can still have pacing, visual relevance or quality problems that automated execution alone will not identify.
His reported measurements also show how subjective feedback can be translated into criteria a production system can test. The figures apply to the first three scenes of this particular video, not to AI video systems generally, and Meyer does not report an independent evaluation. Still, using identified defects to create checks may help prevent the same defects from recurring in later builds. The article’s larger point is that production reliability depends on the process around a model as well as the model’s output.
Products Worth Considering
Apple MacBook Air 13-inch Laptop with M5 Chip, 16GB Memory, 512GB SSD, 12MP Center Stage Camera, Touch ID, Wi-Fi 7, Midnight
Ledger Flex Crypto Wallet with 2.8-Inch E Ink Touchscreen
Logitech MX Master 4 Ergonomic Wireless Mouse with Haptic Feedback, Graphite
As an affiliate, we earn on qualifying purchases.
From One-Minute Test to Full Video
Meyer says the one-minute test video was approved and published the evening the engine was built. The first full video followed the next morning and remains online unchanged as an example of what he considers a poor result. After documenting its shortcomings, he reworked the engine and says a video using the improved look went live a few hours later, in an article about the Nasdaq’s worst day in 19 months. The source does not provide a date for that market event.
The account also links the video work to an unrelated experiment in model editing. Meyer says a model completed 12 of 15 answers in a legal benchmark after a weight-editing experiment, compared with 4 of 15 beforehand, while its legal score remained zero. He uses this result to make a related point: more completed answers did not mean better legal performance. The figures are his account of an in-house experiment, not a claim that the technique improves legal AI generally.
“A demo that runs is not a decision.”
— Thorsten Meyer
Limits of the Reported Measurements
The reported video metrics cover only the first three scenes of the rebuilt video. Meyer does not publish the measurement method, show results across a larger set of videos or provide an external assessment of whether the new version performed better with viewers. The account also does not specify how often the new checks block a render or how reliably they identify irrelevant visual elements.
The source is a first-person account by the person who built the system. Its descriptions of Codex’s findings, the engine’s changes and the model experiment are Meyer’s own reporting; independent verification is not provided. It is also unclear whether the production gates will remain effective as the engine, video subjects or formats change.
Whether the Checks Hold Up
Meyer says the frame-coverage and visual-relevance checks are now part of the build process, so subsequent scenes can be flagged before rendering. The immediate next test is whether those gates continue to catch problems in future videos without blocking acceptable work. The account does not announce a release schedule, a broader performance study or plans to make the engine available to other users.
Readers can compare the published videos and judge the visual changes for themselves, but the source provides no audience data or independent review. Any further claims about quality, speed or reliability would need results from additional videos and a clearer account of how the checks are measured.
Key Questions
What did Meyer build?
He says he built an engine for hand-drawn-style explainer videos, using Claude to write the code and a design system and licensed handwriting font for visual elements.
What was wrong with the first full video?
Meyer says it had too much empty space, icons that did not support the narration and images that stopped moving after they were drawn.
How did he measure the revision?
For the first three scenes, Meyer 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 source does not describe an independent measurement or evaluation.
What roles did the AI tools have?
Meyer says Claude built the engine, while OpenAI’s Codex reviewed the code and reported issues. He says each of five initial findings received a test intended to fail on the old code.
Does the account show that the engine works reliably in general?
No. It describes one creator’s workflow and measurements from a limited set of scenes. The article does not provide broader testing, independent verification or audience results.
Source: ThorstenMeyerAI.com
