🔍 Read the full analysis: An Evening Spent Building An AI Video—and The Lessons It Left Me 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 he built a hand-drawn AI video engine in one evening, but its first full video was slow and visually static. He measured the problems, improved the video and turned the measurements into checks that can stop similar scenes from being rendered.
Thorsten Meyer published an account on November 6 of building a hand-drawn AI video engine in one evening, then revising it after its first full video looked slow and empty. The personal report describes how independent code review, human approval and scene-level measurements shaped the version later published on his site; it is not an independently verified evaluation of the tools or process.
Meyer says the idea came from a hand-drawn explainer whose style he liked. To avoid copying its work, he says he wrote down a limited set of broad visual traits: hand-drawn strokes, a continuous world rather than slides, and large numbers used as color accents. He says the project used no drawings, assets or layouts from the reference video, and that its colors came from his own design system and its handwriting font was openly licensed.
According to Meyer, Claude built the engine, while OpenAI’s Codex reviewed the code without building it. Codex raised five findings in the first review round; Meyer says he fixed each one that evening and added a test that would fail against the old code. He also says he watched a roughly one-minute test video and approved it for publication. That short version appeared in an article about local votes against data centers.
The first full video, made the next morning, did not meet his expectations. Meyer describes too much white space, icons that did not support the narration and completed drawings that remained still. After measuring the first three scenes, he says the rebuilt version raised the share of the frame filled with content from 15% to 66%, reduced stationary time from 29 seconds to zero and increased camera moves from three to 27. These are his project measurements for those scenes, not general benchmarks for AI 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
From Visual Critique to Build Checks
The account is a practical example of a gap between a working demonstration and a video ready for publication. Meyer’s first short test won his approval, but the next morning’s full video exposed problems that had not been apparent in the demo. His response was to translate subjective criticism—too much empty space and too much stillness—into measurements that could be checked during production.
He says those measurements now act as gates: a scene judged too empty can be stopped before rendering, while an object that does not illustrate the narration triggers a warning. That approach matters to readers using generative tools because it makes a human review and explicit quality criteria part of the workflow, rather than treating a successful render or a green test suite as proof that the result is good.
The evidence remains limited to Meyer’s own project notes and experiments. The reported changes cover the first three scenes, and the article does not provide an independent comparison of the complete videos. The account is best read as a documented production lesson, not as proof that the same checks will improve every AI-generated video.
Products Worth Considering
Kindle Paperwhite Signature Edition 32GB (2024) – Auto-adjusting front light, wireless charging, weeks of battery life – Metallic Black
Apple MacBook Air 15-inch with M5 chip, 16GB memory, 512GB SSD, and Liquid Retina display (Midnight)
Apple iPad mini (A17 Pro), 128GB, Wi-Fi 6E, Space Gray
As an affiliate, we earn on qualifying purchases.
How the Video Was Built
Meyer frames the project as a personal account of taking an AI idea into production, including failures. He says the initial visual reference supplied only broad stylistic inspiration, while the engine, artwork and layout were made separately. He identifies Claude as the builder and OpenAI’s Codex as the reviewer, assigning the two systems different roles.
His account also describes other experiments as reminders that completion and passing tests do not necessarily mean a system works as intended. In a separate legal-model experiment, he says an edited model completed 12 of 15 answers on a legal benchmark, compared with four before the edit, while its legal score remained zero. He also reports a review round in which all 12 of his own tests passed but four of a reviewer’s six checks failed. These examples are his reported results, not independent findings about model performance.
Meyer further recalls that a guard produced about 1,300 warnings in August, but nobody read the log and an alert email did not go out. He uses the episode to argue that monitoring must reach a person to be useful. This history helps explain why the video workflow includes both automated checks and a human decision to publish.
“And honestly, it was slow and boring.”
— Thorsten Meyer, describing his first full video
Limits of the Reported Measurements
The measurements and development history come from Meyer’s own build notes, incident notes and experiments. The account does not provide independent verification, a detailed description of how frame fill and stationary time were calculated, or results across a larger set of scenes. It is also unclear how consistently the new gates identify poor scenes or how often they produce false warnings.
The article describes the improved video as live, but does not report audience response or a controlled comparison of viewer engagement. The legal-model example is separate from the video project, and the account does not establish that the weight edit caused a general improvement in legal reasoning. The source material also does not set out a broader assessment of copyright or style-imitation questions beyond Meyer’s description of the steps he took.
The Checks Face Future Videos
Meyer says the scene measurements have been added as production gates, so future renders can be stopped or flagged when they fail his criteria. The next practical test is whether those checks catch problems reliably across additional scenes and videos, while leaving room for a human to judge qualities that numbers may not capture.
The improved video and the original full version remain points of comparison in his account: he says the first full video is still online unchanged as an example of what did not work. No external audit, wider benchmark or later update is included in the material provided. The results beyond Meyer’s own published work remain to be seen.
Key Questions
What did Thorsten Meyer build?
He says he built an engine for hand-drawn explainer videos, with a continuous visual world, animated drawings and camera movement tied to the narration.
Why did the first full video need revision?
Meyer says it had too much white space, icons that did not support the spoken material and drawings that stopped moving after they were finished.
What changed in the rebuilt video?
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 rose from three to 27. He says the figures became checks in the production process.
Which AI systems did Meyer use?
Meyer says Claude built the engine and OpenAI’s Codex reviewed the code. He describes Codex as a reviewer, not a builder, and says he made the final decision about whether a video could go live.
Are the reported results independently verified?
No independent verification is described in the source material. The measurements and workflow details are presented by Meyer as results from his own project notes and experiments.
Source: ThorstenMeyerAI.com
