🔍 Read the full analysis: One Evening To Build With AI: The Lessons Behind The First Cut 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 video engine in one evening, but its first full video was slow and visually weak. After measuring specific problems, he added checks to the production process; a separate model-editing experiment also showed that completing more benchmark tasks did not improve the model’s legal score.
Thorsten Meyer says he built a hand-drawn video engine in one evening, but the first full video it produced the next morning was, in his assessment, slow and boring. In an account published on ThorstenMeyerAI.com and updated November 9, he describes how he revised the system using measured visual checks, a second AI model for code review and a human approval step before publication.
Meyer says the initial idea came from a hand-drawn explainer video he admired. To avoid copying it, he says he wrote down a short list of general style traits, including hand-drawn strokes, a continuous visual world and large numbers used as colour accents. He says the new engine was built from scratch, with no drawings, layouts, assets or drawing code taken from the original. The colours came from his own design system, and the handwriting font was openly licensed.
In his workflow, Claude built the engine and OpenAI’s Codex reviewed the code without building it. Meyer says Codex identified five issues in the first review, all fixed that evening. For each fix, he says, the team added a test that failed against the old code. He describes this as a way to check that a test can detect the original problem, rather than merely pass after a change.
Meyer says he watched and approved a roughly one-minute test video that evening, which was published in an article about local votes against data centers. The first full video in the new style, produced the next morning, drew a different reaction: he found too much empty space, icons that did not support the narration and images that stopped moving once drawn. He says that version remains online unchanged as an example of what did not work.
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
Turning Video Criticism Into Checks
The account highlights a practical distinction for teams using generative AI: a system can produce a finished file without producing a result that meets a human or editorial standard. Meyer says he moved beyond subjective revisions by tracking frame coverage, still time and camera movement across the first three scenes of the revised video.
In those scenes, the filled frame increased from 15% to 66%, still time fell from 29 seconds to zero, and camera moves rose from three to 27. These figures describe the first three scenes, not a general benchmark or independently verified evaluation. Meyer says the measurements were then turned into gates: checks can stop scenes that are too empty, while a build warning flags objects that do not illustrate what is being said.
That approach matters because it links feedback to repeatable production rules. A human still has to judge whether the video is useful and appropriate, but quantitative checks can catch some recurring problems before a render is complete. The account is a first-person report of one creator’s process, not evidence that these particular thresholds will work for every video or production team.
Products Worth Considering
Bose QuietComfort Earbuds (2nd Gen) – True Wireless Bluetooth Earbuds with Active Noise Cancellation
Samsung Galaxy Tab A11+ 8GB RAM, 256GB Storage, Gray
Logitech MX Ergo S Advanced Wireless Trackball Mouse USB-C – Graphite
As an affiliate, we earn on qualifying purchases.
Two Tests of AI Quality
Meyer frames the video project alongside a separate experiment from the same week. He says he asked Claude to edit the weights of a local model in an effort to improve its performance on legal tasks. According to his account, the edited model answered all 10 published tasks in the Lexassay benchmark, compared with four before the edit. The legal score, however, remained at zero.
The contrast is central to Meyer’s stated lesson: a higher completion count does not, by itself, show that answers are correct. The article’s November 9 update says the figures now refer to 10 published Lexassay tasks and that the runs are public on lexassay.com. The source describes Meyer’s own experiments and build notes; it does not provide an independent assessment of the video system or the legal benchmark results.
Meyer also cites an earlier monitoring failure from August. He says a guard generated about 1,300 warnings in a log that nobody read, and an alert email was not sent. He uses that incident to argue that monitoring only helps when it reaches someone able to act. The example puts operational follow-through alongside model output and testing as part of his account of taking AI-generated work into production.
“A demo that runs is not a decision.”
— Thorsten Meyer
Limits of the Published Results
The account does not specify the exact scoring method behind the legal benchmark’s zero score, explain what counted as a correct answer, or provide enough detail to assess the model edit independently. Meyer says the runs are public, but the source material does not establish whether they have been reviewed by external researchers.
For the video work, the reported measurements cover only the first three scenes of the revised video. The account does not give the full-video results, the precise thresholds used by the gates, or an independent evaluation of the finished videos. It also does not describe whether viewers beyond Meyer assessed the revised work. His description of the first video as boring and the improvements as better reflects his own judgment.
The article’s account of the August monitoring problem does not state whether the alert failure was caused by a configuration issue, a delivery problem or another fault. It reports that warnings were logged and an email did not go out, but does not provide a full incident analysis.
Public Runs and Production Checks
Meyer says the 10 Lexassay runs are public on lexassay.com, allowing readers to inspect the benchmark results he cites. The source does not announce a further benchmark date, a new video release schedule or an external review of either experiment.
For the video engine, the next step described in the account is continued use of the checks introduced after the first full video: stop scenes that fail the visual rules, flag mismatches between illustrations and narration, and retain human review before publication. Whether those gates will produce consistently stronger videos across future projects remains to be seen.
Key Questions
What did Meyer build in one evening?
He says he built a hand-drawn video engine inspired by broad stylistic traits from an explainer video. He says the engine used original drawings and code rather than assets or layouts from the video that inspired the idea.
What changed after the first full video?
Meyer says he measured frame coverage, still time and camera movement, then added checks to flag or stop scenes with problems. Across the first three scenes he cites, frame coverage rose from 15% to 66%, still time fell from 29 seconds to zero, and camera moves increased from three to 27.
Did the legal model experiment improve the model?
Meyer says the edited model answered 10 published tasks instead of four, while its legal score remained zero. The higher task-completion count alone does not show that the answers were correct.
Who reviewed the video engine’s code?
Meyer says Claude built the engine and OpenAI’s Codex reviewed the code. He says Codex identified five findings in the first review, which were fixed with tests that failed against the previous code.
Are the results independently verified?
The source presents Meyer’s own account and says the 10 Lexassay runs are public. It does not report an independent evaluation of the video measurements or the legal-model results.
Source: ThorstenMeyerAI.com
