🔍 Read the full analysis: A Night Of Building: Taking An AI Idea Toward Production on ThorstenMeyerAI.com
Get the little things that make your day 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 and publishing a hand-drawn video engine in a few hours, then revising it after the first full video proved slow and visually weak. The account says the workflow now includes review by OpenAI’s Codex, human sign-off and measurable checks that can block scenes. It also reports a separate legal-model experiment in which task completion rose while the legal score remained zero.
Thorsten Meyer says he built and published a hand-drawn video engine within hours, then changed its production process after the first full video came out slow and visually sparse. In an account published on ThorstenMeyerAI.com and updated November 9, he describes adding independent AI code review, human approval and automated quality checks—a case study in why a working AI demo is not, by itself, evidence of a production-ready result.
Meyer said the idea came from a hand-drawn explainer video whose style he admired. To avoid copying it, he said he used a “clean room” approach: recording a short list of general traits, including hand-drawn lines and a continuous visual world, while using no drawings, assets, layouts or drawing code from the original. He said Claude built the engine from scratch, using his own design system and an openly licensed handwriting font.
For review, Meyer said he assigned OpenAI’s Codex a separate role: it checks code but does not build. Codex identified five issues in the first review round, he wrote, and each fix was paired with a test designed to fail on the old code. Meyer argues this helps show that a test can detect the original problem rather than merely confirm a later change. He also describes his own approval as a required step: after watching a roughly one-minute test video, he approved it for publication.
The first full video made the next morning did not meet his expectations. Meyer cited excessive white space, icons that did not support the narration and images that stopped moving after they were drawn. He said the video remains online unchanged as an example of what went wrong. The team then measured frame coverage and still-image time and added checks to the build. In the first three scenes of the rebuilt video, he reported, frame coverage rose from 15% to 66%, standing still fell from 29 seconds to zero, and camera moves increased from 3 to 27. The article says those figures now inform gates that can stop a scene if it is too empty or an object does not match 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 Demo to Quality Gates
The account highlights a practical gap in AI production: a system can run and produce a finished file without producing something useful or accurate. Meyer’s response was to turn subjective feedback into measurable checks and make those checks part of the build, rather than relying on repeated manual tweaks. That approach may help teams catch known presentation problems before publication, although the reported measurements cover only the first three scenes and do not by themselves establish the quality of every video.
The workflow also divides responsibility among tools and people. Claude generated the engine, Codex reviewed code, and Meyer made the publication decision. That distinction matters because automated review and tests can catch particular errors, but the account does not treat either as a substitute for human judgment. Meyer’s description is a personal report, not an independently verified evaluation of the system or a controlled comparison with other production methods.
A separate experiment in the same account illustrates the risks of choosing an incomplete success metric. Meyer says a model edited by Claude answered all 10 published legal benchmark tasks, up from four, while its legal score remained at zero. The result underscores that completion rate and task quality are different measures; a better-looking headline number can coexist with no improvement on the score the experiment was meant to improve.
Products Worth Considering
Bowflex SelectTech 840 Adjustable Kettlebell
Philips Sonicare ProtectiveClean 5300 Rechargeable Electric Toothbrush with 3 Brush Heads, Black
Hyperice Normatec Elite Wireless Compression Boots
As an affiliate, we earn on qualifying purchases.
How the Video Workflow Evolved
Meyer dates the initial idea and first test to one evening. He said the roughly one-minute version went live that night in an article about local votes against data centers. The first full-length production followed the next morning and prompted a written critique. A rebuilt version, he said, was published a few hours after that critique, in an article about the Nasdaq’s worst day in 19 months.
The November 9 update says figures in the article now refer to 10 published Lexassay tasks and that experiment runs are public on lexassay.com. Meyer presents both the video engine and the legal-model test as examples from his own work, drawing on his build notes, incident notes and published experiments. The source does not provide an outside audit of the results.
The account also describes a monitoring failure from August: a guard generated about 1,300 warnings in a log no one read, while an alert email did not go out. Meyer uses that incident to argue that monitoring only serves its purpose when a warning reaches a person who can act. The article does not specify the system involved or provide further details about the missed email.
“A demo that runs is not a decision.”
— Thorsten Meyer
Limits of the Reported Results
The reported video measurements cover the first three scenes of the rebuilt production; the account does not provide equivalent figures for the full video or a broader set of productions. It also does not describe how frame coverage, stillness or whether an object illustrates the narration are calculated, or how often the new gates block a build.
The article is Meyer’s own account of his work. It does not include independent testing of the engine, external review of the legal benchmark results, or details that would allow readers to reproduce every experiment. Although Meyer says the legal model answered all 10 published tasks while retaining a score of zero, the supplied account does not explain the scoring method or why the score did not change.
Further Runs and Production Checks
Meyer says runs for the 10 published Lexassay tasks are available on lexassay.com, giving readers a place to inspect the legal experiment’s results. The article does not announce a schedule for additional tests or specify further releases of the video engine.
For the video workflow, the next stated step is continued use of the new build gates: scenes that are too empty are stopped, and objects that do not illustrate the narration prompt a warning. Whether those checks improve results consistently across later videos remains to be shown. The account’s own standard is clear: publication follows code checks, measurements and a human decision—not simply a successful render.
Key Questions
What did Meyer build?
He says he built a hand-drawn video engine from scratch after defining broad stylistic traits from a video he admired. He says he did not use the original’s drawings, assets, layouts or drawing code.
What changed after the first full video?
After finding too much white space, weakly connected icons and static finished drawings, Meyer says the team measured frame coverage and stillness and added checks to block scenes that fail those criteria.
What do the reported video numbers cover?
The comparison is for the first three scenes of the rebuilt video: coverage rose from 15% to 66%, stillness fell from 29 seconds to zero, and camera moves rose from 3 to 27. The article does not give those measurements for the entire video.
Did the legal-model experiment improve the model’s score?
Not according to Meyer’s account. The model answered all 10 published tasks, compared with four before the edit, but its legal score remained zero. The account does not explain the scoring method.
Who approved the video for publication?
Meyer says he watched a roughly one-minute test version and personally approved it. He describes Codex as a code reviewer, not the tool responsible for making the video or deciding whether it should go live.
Source: ThorstenMeyerAI.com
