🔍 Read the full analysis: From AI Idea To Production: Lessons From One Evening And A First Video on ThorstenMeyerAI.com
Get movie nights 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 video engine in one evening, but its first full video was slow and visually sparse. After measuring specific flaws, he rebuilt the engine and turned those measurements into checks that can stop a weak scene before rendering. The account also describes independent AI code review, human approval and separate experiments that show why passing a basic completion test does not prove quality.
Thorsten Meyer says he built a hand-drawn video engine in one evening, then found its first full video too slow and visually dull the next morning. In an account published on ThorstenMeyerAI.com and updated November 9, Meyer describes how he measured the problems, revised the engine and added automated checks—an example, in his telling, of why a working AI demo still needs review before it is treated as production-ready.
Meyer says the idea came from a video style he admired: a hand-drawn explainer moving through a continuous whiteboard-like world. He describes setting a clean-room boundary before building: his team wrote down a short list of broad visual traits, but did not reuse the original video’s drawings, assets, layouts or code. The new engine used his own design system, including blue for the main imagery and orange accents, along with an openly licensed handwriting font.
According to Meyer, Claude built the engine, while OpenAI’s Codex reviewed the code without contributing to the build. He says Codex identified five issues in its first review, and that each was fixed alongside a test designed to fail on the old code. Meyer argues that this sequence matters because a test that fails before a fix offers evidence it can detect the original problem, rather than merely confirming the revised code.
Meyer says he watched and approved a test video of about one minute that evening; it was published within an article about local votes against data centers. The first full video, produced the next morning, did not meet his expectations. He left it online unchanged as an example of the initial result, describing excessive white space, icons that did not support the narration and drawings that stopped moving once completed.
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
How Video Critique Became a Build Gate
The account’s central point is that a fast first draft is not the same as a reliable production process. Meyer says his team turned subjective complaints about the video into measurements, then used those measurements to block some failures before a scene is rendered. On the first three scenes of the rebuilt video, the reported share of the frame filled with content rose from 15% to 66%; time spent completely still fell from 29 seconds to zero; and camera moves increased from 3 to 27. These figures apply to those three scenes, not necessarily to every video or the whole production process.
Meyer says the new checks stop scenes that are too empty and flag objects that do not illustrate what is being said at that moment. That approach matters to teams using generative tools because it makes some standards repeatable: a reviewer can identify a problem, and a test can help prevent the same specific problem from returning. The safeguards described are limited to the criteria he chose; they do not establish that the video is accurate, effective or suitable for every viewer.
His separate legal-model experiment reinforces the distinction between a visible output and a meaningful result. Meyer reports that after a weight-edit experiment, a model answered 10 published legal tasks, compared with four before, while its legal score remained at zero. The source does not provide enough detail here to independently assess the benchmark or explain the score, but the reported result illustrates why task completion alone can give a misleading impression of performance.
Popular right now
Apple AirPods Pro (2nd Generation) Wireless Earbuds with USB‑C Charging, Up to 2X More Active Noise Cancelling, Transparency Mode, Adaptive
Vitamix Touch & Go Advance 32-Ounce Bar Blender, White
Bose QuietComfort Earbuds (2nd Gen) – True Wireless Bluetooth Earbuds with Active Noise Cancellation
As an affiliate, we earn on qualifying purchases.
From Style Reference to Revised Video
Meyer dates the original account November 6, 2026, and says it was updated November 9. The update states that figures now refer to 10 published Lexassay tasks and that the runs are public on lexassay.com. His narrative moves from an initial evening prototype to a test video approved for publication, then to a disappointing full video the following morning and a revised version published a few hours after his critique.
The production process he describes has several distinct checks: one AI creates the engine, a second AI reviews code, tests are written to catch identified defects, and a person responsible for the result decides whether a video is ready to publish. The later measurements and build gates address particular visual and narrative issues. Meyer also recounts an earlier monitoring failure from August: a guard generated about 1,300 warnings in a log nobody read, while an alert email did not go out. That incident led him to stress that monitoring only helps when its output reaches a person who can act on it.
“A demo that runs is not a decision.”
— Thorsten Meyer
What the Reported Metrics Establish
The figures come from Meyer’s own build notes, incident notes and experiments, as described in his article; they are not presented as an independent evaluation. The reported video measurements cover the first three scenes of the rebuilt video, and the account does not give a full-video comparison or explain how each metric was calculated. It is also unclear from the supplied material how the team evaluated whether the revised video better communicated its subject to viewers.
The source describes the engine’s checks and says a stronger-looking video went live, but it does not establish that the safeguards catch every visual, factual or editorial problem. The legal-task result is similarly limited: the article reports 10 answers rather than four and a legal score of zero, while the excerpt does not state the scoring method or provide the benchmark’s underlying results. The August warning count also has no comparison baseline, so it should be read as a raw count, not evidence of a particular rate or increase.
Further Tests and Public Run Records
Meyer says the legal benchmark runs are public on lexassay.com, following the November 9 update, so readers can consult the published tasks and runs. The account does not announce a date for another release or a new evaluation. The next useful evidence would be results across complete videos and additional scenes, alongside clear descriptions of how frame coverage, stillness, narrative relevance and legal scores are measured.
For now, the reported process is an account of one builder’s workflow rather than a controlled comparison of AI production methods. Meyer says he will keep the first unsatisfactory video online as a reference point; whether the new gates consistently prevent similar problems in later work remains to be shown.
Key Questions
What did Thorsten Meyer build?
Meyer says he built an engine for hand-drawn-style explainer videos in one evening, using a continuous visual world rather than a sequence of conventional slides.
What was wrong with the first full video?
Meyer described it as slow and boring, with too much white space, icons that did not support the narration and drawings that became still after they were completed. He says that version remains online unchanged.
How did the team measure the revised video?
For the first three scenes of the rebuilt video, Meyer reports frame coverage increasing from 15% to 66%, stillness falling from 29 seconds to zero, and camera moves rising from three to 27. These are reported measurements for those scenes, not a general result for all videos.
Did an AI approve the video for publication?
No. Meyer says Codex reviewed the code, while a human watched and approved the test video. He describes the review and the human publication decision as separate steps.
What remains unverified?
The account does not provide an independent assessment of the reported metrics, a full-video comparison or evidence that the gates catch every kind of problem. The legal benchmark’s scoring details are also not included in the supplied material.
Source: ThorstenMeyerAI.com
