🔍 Read the full analysis: What I’d Do Differently Taking An AI Idea To Production on ThorstenMeyerAI.com
Get the little things that make your day 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 idea became a live format within hours, but the first full video exposed problems that a quick demo had not. His account describes using independent AI code review, human approval, measurements and automated gates to improve the production process; it is a personal report, not an independently verified study.
Thorsten Meyer says he turned an idea for a hand-drawn, whiteboard-style video into a live format within hours, then rebuilt the approach after the first full video proved slow and visually dull. In a personal account published November 6 on ThorstenMeyerAI.com, he describes adding independent AI code review, human approval and automated quality checks—a process he presents as a way to distinguish a working demo from a production-ready result.
Meyer says the idea came after watching a video whose visual style he liked. To avoid copying it, he says he wrote down a short list of general traits—such as hand-drawn lines and a continuous illustrated world—and did not use the original video’s drawings, assets or layouts. He says Claude built a new drawing engine, while colors came from his own design system and the handwriting font was openly licensed.
In his workflow, OpenAI’s Codex reviews rather than builds. Meyer says Codex identified five issues in its first review; he reports fixing each one and adding a test that would fail against the earlier code. He also describes watching a roughly one-minute test video himself before approving it for publication. The initial short version went live inside an article about local votes against data centers, according to his account.
The first full video produced the following morning did not meet his expectations. Meyer says it had too much empty space, icons that did not support the narration and images that stopped moving after they were drawn. He measured frame coverage and still-image time, then says he revised the engine. Across the first three scenes, the filled portion of the frame rose from 15% to 66%, still time fell from 29 seconds to zero, and camera moves increased from 3 to 27. These are figures for those scenes in his own project, not results from an external evaluation.
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 Demo to Production Checks
The account highlights a practical problem for teams adopting AI tools: a prototype can run and still produce work that is not useful or ready to publish. Meyer’s example shows how a subjective reaction—his description of the first full video as boring—was turned into specific production checks, including frame coverage and motion. He says those measurements now act as gates that can stop a scene before rendering.
His distinction between a test passing and a result being good also matters beyond video. In another experiment described in the article, a model completed 12 of 15 legal-benchmark answers, compared with four previously, while its legal score remained zero. The result, as Meyer reports it, illustrates why completion counts alone can give a misleading impression of quality. The account does not establish how the benchmark was designed or whether the result generalizes to other models or tasks.
Meyer also points to a monitoring failure from August: he says a guard produced about 1,300 warnings in a log nobody read, while an alert email did not go out. The example underscores that logging is not effective oversight unless someone receives and acts on the signal. Taken together, the cases make the article relevant to developers and publishers weighing how to review, test and monitor AI-assisted work.
Products Worth Considering
Sonos Era 100 SL Compact Microphone-Free Speaker, White
Arlo Essential Outdoor Pan-Tilt-Zoom Security Camera, 2K, Wired, 2-Camera Pack
Dyson Cyclone V10 Origin Cordless Vacuum
As an affiliate, we earn on qualifying purchases.
How the Video Workflow Changed
The sequence Meyer describes began with a style idea, followed by an evening prototype and a short test video. He says he approved that version after watching it, and it went live that night. The next morning’s full video exposed problems that the short test had not: excessive white space, weak links between images and narration, and static scenes.
Instead of relying only on repeated visual tweaks, Meyer says he measured two features and rebuilt the engine around them. The resulting rules require fuller frames, camera movement linked to narration and continuing subtle motion in completed drawings. He says objects must also illustrate something being said at that moment. The numerical thresholds are now checks in the build process, although the article does not give every threshold or describe how all visual judgments are automated.
The source is Meyer’s first-person account of his own build notes, incident notes and experiments. It links to a roughly four-and-a-half-minute explainer made using the process and says the narration uses an AI voice clone of Meyer. The article’s figures and descriptions should be understood as claims about his project rather than independently audited findings.
“A demo that runs is not a decision.”
— Thorsten Meyer
Limits of the Reported Results
The article does not provide an independent review of the video metrics, the code-review findings or the legal benchmark. It also does not specify the benchmark’s scoring method, the model’s earlier legal score beyond saying it remained at zero, or whether the 12 completed answers were accurate. The 15-to-66% frame-coverage result and other video figures apply to the first three scenes, not necessarily the whole video or later projects.
It is also unclear how much of the new quality gate is automated, how its thresholds were selected, and whether the checks catch problems such as an image that technically matches narration but communicates it poorly. Meyer says the original full video remains online unchanged as an example of what went wrong. The source does not provide audience-response data or evidence that the revised format performs better with viewers.
The Checks in Future Videos
Meyer says the frame and motion measurements are now gates in the production process: a scene that is too empty can be stopped before rendering, and a mismatch between an image and the narration prompts a build warning. The published explainer is presented as another video made with the revised method. The article does not announce a release schedule, broader rollout or a formal evaluation plan.
The next useful evidence would be whether the checks continue to catch defects across more scenes and projects, whether human reviewers still find problems the automated rules miss, and whether the monitoring alerts reach someone able to respond. Those outcomes are not established in the account. For now, the reported development is a specific creator’s evolving workflow, rather than proof that the same safeguards will produce the same results elsewhere.
Key Questions
What did Thorsten Meyer publish?
He published a first-person account of developing a hand-drawn video engine and changing its production process after the first full video fell short of his expectations.
What changed in the revised video workflow?
Meyer says he measured frame coverage and still-image time, then added build checks for empty scenes and illustrations that do not match the narration. He also describes using code review and human approval.
What do the video metrics cover?
The reported changes—from 15% to 66% frame coverage, 29 seconds to zero of still time, and 3 to 27 camera moves—apply to the first three scenes of the rebuilt video, according to Meyer.
Does the article prove the process works for other teams?
No. It reports Meyer’s experience and project figures. The source does not provide independent validation or evidence that the workflow will produce the same results in other settings.
What remains unknown about the legal-model experiment?
Meyer reports that the model completed 12 of 15 answers rather than four, while its legal score remained zero. The article does not explain the benchmark’s scoring method, accuracy results or whether the change generalizes.
Source: ThorstenMeyerAI.com
