🔍 Read the full analysis: The Hard-Won Rules I Learned Taking An AI Idea To Production on ThorstenMeyerAI.com
Get smart everyday buys 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 a hand-drawn video engine went from an evening idea to a live test video after a second AI reviewed its code and he approved the result. A disappointing full-length video led to measurable production gates, while separate experiments showed that passing tests and completing more answers do not by themselves prove an AI system works well.
Thorsten Meyer says he took a new AI-generated video style from an evening prototype to a live test video, then changed the production process after the first full video proved slow and visually dull. In an account published November 6, he describes using a second AI to review code, requiring his own approval and turning video-quality measurements into checks that can stop later builds.
Meyer said he drew inspiration from a hand-drawn explainer video but set a clean-room rule for his project: record broad style traits, such as a continuous illustrated world and hand-drawn strokes, without reusing the original video’s drawings, assets or layouts. He said Claude built a new drawing engine from scratch, using his own design system and an openly licensed handwriting font.
He assigned OpenAI’s Codex a separate role: review, not construction. Meyer said Codex found five issues in its first review, and he fixed each with a test designed to fail against the earlier code. He argues that this makes the test evidence that it can detect the problem, rather than a check written only to pass after the repair.
Meyer watched a roughly one-minute test video and approved it for publication that evening. The next morning’s first full video drew a different response: he described it as boring, citing empty space, icons that did not support the narration and images that stopped moving when their drawings were complete. He said that version remains online as an example of what the team wanted to avoid.
Rather than adjust the video solely by feel, Meyer said he measured frame coverage and time spent completely still. Across the first three scenes of the rebuilt video, filled-frame coverage rose from 15% to 66%, still time fell from 29 seconds to zero, and camera moves rose from three to 27. These are figures from those scenes, not a general benchmark. He said the measurements now act as build gates: scenes that are too empty are stopped, and objects that do not illustrate the narration trigger a warning.
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 Video Metrics to Production Gates
The account matters because it separates a working AI demonstration from a reviewed and accepted production result. Meyer describes several distinct checks: code review by a second model, tests for identified defects, a human decision about whether the output is publishable, and automated checks against measurable presentation standards.
His examples also show why a single success measure can mislead. In a separate experiment involving a locally run model and legal answers, Meyer said the model completed 12 of 15 answers, up from four, but its legal score remained at zero. The higher completion count did not establish better legal performance. He also reported that in one code-review round all 12 of his own tests passed while four of six reviewer checks failed. Both examples are his reported experiments, not independently verified evaluations.
For readers building AI products, the practical point is that output volume, green tests and a convincing demo each answer different questions. They do not, on their own, establish that a system is accurate, useful or ready to publish. Meyer’s approach puts a human decision and checks tied to observed failures between a prototype and release.
Products Worth Considering
LG UltraGear 32GS60QC-B 32-inch Curved QHD Gaming Monitor (180Hz, 1ms, 1000R, AMD FreeSync, HDR10)
Dell UltraSharp 27 Thunderbolt Hub Monitor (U2726DER), QHD
LG 27UP850K-W 27-inch 4K UHD IPS Monitor
As an affiliate, we earn on qualifying purchases.
How the First Video Was Rebuilt
Meyer’s account is a first-person description on ThorstenMeyerAI.com, not an independent audit of the tools or results. He said the style idea came from watching another explainer and that his team limited itself to general traits rather than copying its visual materials. The original one-minute test video appeared in an article about local votes against data centers.
The first full video exposed problems that the shorter test had not. Meyer said the team then redesigned the engine so shots fill more of the frame, the camera moves with the narration, finished drawings retain slight motion, and each object relates to what is being said. He said the first video using the improved look went live a few hours after his critique, in an article about the Nasdaq’s worst day in 19 months. The source also links a roughly four-and-a-half-minute explainer about the process.
Meyer connects the video work to a broader lesson from his experiments. He said a guard had written about 1,300 warnings to an unread log in August, while its alert email did not go out. The figure is a raw count from his notes; the account gives no comparison period or baseline. It illustrates a separate operational issue: logging a warning does not mean a person has been alerted.
““A demo that runs is not a decision.””
— Thorsten Meyer
What the Report Does Not Establish
The measurements and experiment results come from Meyer’s own build notes and reports. The article does not provide the underlying datasets, full test definitions or independent evaluations, so readers cannot assess how the results would hold across other videos, models or tasks. The frame-coverage figures apply to the first three scenes of one rebuilt video.
The source does not specify the legal benchmark’s full methodology, what the zero score measures, or why the model’s completion rate rose. It also does not say whether the alert-email failure was later corrected or whether the video gates have been tested across a larger body of work. Those details remain unclear from the account.
Checks for Future Builds
Meyer says the frame-coverage and motion checks now form part of the video build process, with rules intended to halt scenes that are too empty or flag visuals that do not match the narration. The immediate next step described is continued use of those gates on new videos; the account does not give a schedule for further releases or a plan for publishing a broader evaluation.
For the separate monitoring and model experiments, the source does not report a specific next milestone. Whether the alert route was repaired, whether legal performance improves beyond answer completion, and how consistently the new video checks work in later projects remain open questions.
Key Questions
What did Thorsten Meyer take from idea to production?
He describes building a hand-drawn video engine inspired by broad traits of an explainer style. He said the project used original drawings and layouts rather than copying the source video’s assets.
How did Meyer review the AI-generated code?
He said Claude built the engine and OpenAI’s Codex reviewed it. Meyer reported five findings in the first review, each followed by a test that failed on the earlier code.
What changed after the first full video?
Meyer said he measured frame coverage and still time, then made them production checks. In the first three scenes of the rebuilt video, coverage increased from 15% to 66% and still time fell from 29 seconds to zero.
Did the legal-model experiment show the model had improved?
Not on the reported legal score. Meyer said the model completed 12 of 15 answers rather than four, but its legal score remained zero. The account does not provide enough benchmark detail to assess the result independently.
What remains unverified?
The article presents Meyer’s own measurements and notes. It does not provide independent verification, full benchmark methods or results across a wider set of videos and models.
Source: ThorstenMeyerAI.com
