🔍 Read the full analysis: How A Simple AI Video Helped Me Learn Production Rules 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 building a hand-drawn AI video engine and revising it after the first full video was criticized as slow and empty. Measurements from its first three scenes improved, and those metrics became checks that can stop later builds; the account is a personal report, not an independent evaluation.
Thorsten Meyer says he turned a hand-drawn AI video idea into a working production workflow in a few days, after the first full video exposed problems with pacing and visual density. In an account published on ThorstenMeyerAI.com and updated November 9, he describes adding independent AI code review, human approval and measurable checks intended to prevent similar production flaws from recurring.
Meyer says the initial idea came from a video whose whiteboard-like, hand-drawn style he admired. To avoid reproducing that work, he says he wrote down a limited set of general traits and built a new drawing engine from scratch. The account says the project used Claude to build the engine and OpenAI’s Codex to review its code. Meyer says Codex identified five issues in its first review round; each fix was paired with a test designed to fail against the earlier code.
A short test video of about one minute received Meyer’s approval and went live that evening in an article about local votes against data centers. The first full video made the next morning did not meet his expectations. He described it as slow and boring, citing large areas of unused white space, icons that did not support the narration, and finished drawings that remained still. Meyer says that version remains online unchanged as an example of what the team wanted to avoid.
Instead of relying only on subjective revisions, Meyer says the team measured frame coverage and the duration of still images, then changed the engine. On the first three scenes of the rebuilt video, he reports that the frame filled with content rose from 15% to 66%, still time fell from 29 seconds to zero, and camera moves rose from three to 27. The account says these measures now act as gates: a scene that is too empty can be stopped before rendering, while an object that does not match the narration can 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
How Video Metrics Became Build Gates
The account offers a practical example of a broader production challenge: a system can generate an output without producing something a person considers ready to publish. Meyer’s process links subjective critique to observable checks, then uses those checks to flag issues before another video is rendered. That may help teams catch repeatable defects earlier, though the article does not establish how well the approach performs across other creators, tools or types of video.
The workflow also assigns different roles to people and AI systems. Meyer says one model generated the engine, another reviewed it, and a human watched the video and decided whether it could go live. The distinction matters because code that passes tests, or a video that finishes rendering, does not by itself show that the result is accurate, useful or suitable for publication.
Products Worth Considering
Samsung Odyssey G9 QD-OLED 49 Dual QHD Curved Gaming Monitor
Apple Watch Series 8 GPS 41mm Midnight Aluminum Case with Midnight Sport Band
Bose TV Speaker Compact Soundbar with Bluetooth and HDMI-ARC, Black
As an affiliate, we earn on qualifying purchases.
From Style Reference to Original Engine
Meyer frames the project as a personal account of taking an AI idea into production, including failures. He says the team recorded only broad visual characteristics from the video that inspired him: hand-drawn strokes, a continuous world rather than slides, and large numbers used as color accents. He says it did not use the original video’s drawings, layouts, assets or drawing code. The new engine used his own design system, an openly licensed handwriting font and a palette in which blue carries the image and orange highlights a key point.
The article’s November 9 update says figures now refer to 10 published Lexassay tasks, with runs made public on lexassay.com. Meyer also describes a separate experiment in which a model edited by Claude answered all 10 tasks in a legal benchmark, compared with four before the edit, while its legal score remained zero. He presents this as a warning that a higher completion count is not necessarily evidence of better task performance. These figures are from the author’s own experiments, not an independently reported evaluation in the supplied material.
““A demo that runs is not a decision.””
— Thorsten Meyer
What the Report Does Not Establish
The measurements are reported by Meyer for the first three scenes of one rebuilt video. The article does not provide an independent audit, a larger sample of videos, or evidence that the same gains hold across different topics and production conditions. It also does not specify the exact thresholds used by the new gates or how often they stop a build.
The account describes Codex’s review and the associated tests, but the supplied material does not include the full code, test suite or reviewer findings. It is also unclear how the legal benchmark’s score is calculated, why it remained at zero after the model answered all 10 tasks, or whether those responses were correct. The completion count alone cannot answer those questions.
The Next Test Is Repeated Use
Meyer says the new video engine now uses its measurements as checks before rendering, and that a video about the Nasdaq’s worst day in 19 months was published in the improved style a few hours after his critique of the earlier full video. The updated account also points readers to public runs for the 10 Lexassay tasks. No further release date or independent assessment is stated in the source material.
The next evidence to watch for is whether the gates catch problems reliably across more videos, and whether the resulting work meets human standards beyond the creator’s own approval. For the legal experiment, a useful next step would be reporting task-level correctness and the benchmark’s scoring method, rather than completion totals alone.
Key Questions
What happened in Meyer’s AI video project?
Meyer says he built a hand-drawn video engine, found the first full video too slow and visually sparse, and revised the system using measurements and production checks.
What changed in the rebuilt video’s first three scenes?
According to Meyer, frame coverage rose from 15% to 66%, still time fell from 29 seconds to zero, and camera moves increased from three to 27. Those figures apply to the first three scenes described in his account.
What roles did the AI tools have?
Meyer says Claude built the engine, while OpenAI’s Codex reviewed the code and reported possible issues. He says a human still watched and approved the video before publication.
Do the reported results prove the workflow works generally?
No. The account reports results from Meyer’s own project and experiments. The supplied material does not provide an independent evaluation or evidence from a broad set of videos.
Why does the article mention a legal-model experiment?
Meyer uses it to illustrate the limits of completion counts: a model answered 10 published tasks instead of four, but its legal score remained zero. The source does not provide enough detail to assess the benchmark or the answers’ accuracy.
Source: ThorstenMeyerAI.com
