🔍 Read the full analysis: The Reality Of Taking AI From Idea To Production 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 describes taking a hand-drawn AI video concept from an evening prototype to published work, with revisions driven by measured defects and checks that can stop a weak scene. A separate legal-model experiment and software review findings reinforce the limits of completion rates and green tests. The account is based on the author’s own work and experiments, not an independent evaluation.
Thorsten Meyer has described how he took an AI-generated, hand-drawn video concept from an evening prototype to published videos, using code review, human approval and measurable production checks to address problems that appeared after the first demo. The account, published on ThorstenMeyerAI.com and updated November 9, is a personal report based on his build and incident notes, rather than an independent assessment of the approach.
Meyer said the initial concept came from a video style he liked, but he set a clean-room rule to avoid copying its drawings, assets or layouts. The team wrote down a short list of general style traits, including hand-drawn lines and a continuous visual world, then built a separate engine. Meyer said the color scheme came from his own design system and the handwriting font was openly licensed.
In his workflow, Claude builds the engine and OpenAI’s Codex reviews the code without contributing to its construction. Meyer said Codex reported several findings in its initial review; five were fixed that evening, each with a test designed to fail on the earlier code. He also described watching a roughly one-minute test video himself and approving it before publication. That 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 said it had too much empty space, icons that did not support the narration and images that stopped moving after they were drawn. He kept that video online unchanged as an example of the initial result. The team then measured frame coverage and time spent still. On the first three scenes of the rebuilt video, he reported that filled-frame coverage rose from 15% to 66%, still time fell from 29 seconds to zero, and camera moves increased from three to 27. He said those measures now act as production gates, though the account does not specify their implementation in detail.
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 Release Gates
The report illustrates a practical distinction for teams adopting generative AI: producing a working demo is not the same as deciding that its output is suitable to publish. Meyer’s account places human review alongside automated checks, rather than treating either as a substitute for the other. His approval was a release decision; the subsequent critique identified shortcomings the initial test had not caught.
Turning those shortcomings into measurable checks could make similar problems easier to catch before later videos are rendered. But the figures describe only the first three scenes of one rebuilt video, as reported by Meyer. They do not establish that the same process will improve every video, or that frame coverage and motion alone measure quality. The account is useful as a case study in workflow design, not proof of a general performance claim.
Meyer also reported a separate legal-model experiment that highlights the risk of choosing an incomplete success measure. A model answered more benchmark tasks after a weight-edit experiment, but its legal score remained zero. That result underlines why task completion cannot stand in for correctness when the measure that matters is an answer’s quality.
Products Worth Considering
Logitech MX Ergo S Advanced Wireless Trackball Mouse USB-C – Graphite
Samsung Galaxy Tab A11+ 11-Inch Tablet, 6GB RAM, 128GB Storage, Expandable, Gray
Bose QuietComfort Headphones (2nd Gen), Wireless Over-Ear Noise-Cancelling Headphones, Black
As an affiliate, we earn on qualifying purchases.
How the First Video Fell Short
Meyer’s timeline began with a style idea, followed by an engine built in one evening. A short test video was approved and published that night. The next morning’s full-length production exposed problems that the test had not made apparent, prompting him to document specific criticisms and revise the engine.
The report also describes two earlier or parallel lessons about checking AI work. In a software review, Meyer said all 12 author-written tests passed, while four of six checks from the reviewer failed. In an incident from August, he said a guard produced about 1,300 warnings in a log nobody read, and an alert email did not go out. These are the author’s reported examples; the article does not provide outside verification or enough detail to assess the systems independently.
In the legal experiment, Meyer said the model answered all 10 published Lexassay tasks after its weights were edited, compared with four before, while its legal score stayed at zero. The page was updated November 9 to say the figures refer to those 10 published tasks and that runs are public on lexassay.com. The raw task counts do not by themselves establish a broader improvement in legal performance.
“A demo that runs is not a decision.”
— Thorsten Meyer
What the Account Cannot Establish
The report gives one creator’s account of a video workflow and selected measurements. It does not provide an independent review, a full description of how the measurements were collected, or results across a larger sample of videos. It is also unclear how the new gates define an overly empty scene or decide whether an object accurately illustrates narration.
The claims about code-review findings, test results, the August alert failure and the legal benchmark are attributed to Meyer. The account does not include enough technical detail to independently verify them or compare the workflow with other production systems. The legal task count and score also do not establish how the model performs on legal work outside the published benchmark.
Finally, Meyer’s article does not report whether the process has reduced defects over time, how much it costs to maintain, or whether other people can reproduce the results. Those questions remain open beyond the examples he describes.
Further Runs and Production Checks
Meyer says the legal experiment’s runs are public on lexassay.com, giving readers a way to inspect the benchmark results he cites. His account does not announce a planned external audit or a new release date for the video engine.
The immediate next step in the workflow he describes is continued use of the checks: scenes that fail the stated measures can be stopped before rendering, while people remain responsible for reviewing the finished video. Whether those gates catch more problems in future work, and how reliably they reflect audience-facing quality, will require results beyond the examples in the report.
Key Questions
What is the development being reported?
Thorsten Meyer’s account describes a hand-drawn AI video engine moving from a prototype to published videos, and the review, measurement and approval steps he used to revise it.
What changed after the first full video?
Meyer said the first full video was too empty, visually static and poorly supported by its icons. The team measured frame coverage and still time, revised the engine, and added checks intended to stop scenes that fail its criteria.
Do the reported figures prove that the new workflow works generally?
No. The video measurements concern the first three scenes of one rebuilt video, as reported by Meyer. They do not establish results across other videos or show that the measures capture every aspect of quality.
What happened in the legal-model experiment?
Meyer said a model answered all 10 published tasks after its weights were edited, compared with four before, but its legal score stayed at zero. The reported task count is not evidence by itself that the model became legally reliable.
What remains unknown?
The report does not give independent verification, broad results across multiple productions, or enough technical detail to evaluate how the gates work. It also does not show whether the process reduces defects over time or can be reproduced by other teams.
Source: ThorstenMeyerAI.com
