🔍 Read the full analysis: How I Learned The Rules Of AI Production By Making A Video 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 and published a new hand-drawn video look after testing the first full version, finding it slow and visually static, and revising the production process. His account describes separate AI roles, human approval, measurable scene checks and automated gates; the results and measurements come from his own project notes.
Thorsten Meyer says he turned an idea for a hand-drawn explainer video into a live production style this week, after the first full video exposed problems with empty frames and static images. In an account published November 6 on ThorstenMeyerAI.com, Meyer describes the changes he made and the safeguards he now uses, including independent AI code review, a human approval step and measurable checks that can stop a scene from being rendered.
Meyer says the idea came from watching a video whose whiteboard-like style he liked. To avoid copying it, he says he wrote down a limited set of general visual traits: hand-drawn lines, a continuous world rather than slides and large numbers used as color accents. He says his team used no drawings, layouts, assets or drawing code from the reference. Claude built the engine, while OpenAI’s Codex reviewed the code without building it. Meyer says five findings from Codex’s initial review were fixed that evening, with each fix paired with a test designed to fail on the earlier code.
A roughly one-minute test video received Meyer’s approval and appeared in an article about local votes against data centers. The following morning, however, the first full video in the style struck him as slow and boring. He identified excessive white space, icons that did not support the narration and images that stopped moving after they were drawn. Meyer says that video remains online unchanged as an example of what the team wanted to improve.
Rather than making adjustments only by feel, Meyer says the team measured how much of each frame contained content and how long the image remained completely still. In the first three scenes of the rebuilt video, he reports that frame fill rose from 15% to 66%, still time fell from 29 seconds to zero, and camera moves increased from three to 27. He says those measurements became production gates: checks can stop scenes that are too empty, and the build can flag objects that do not illustrate the words being spoken.
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 Critique to Production Gates
The account matters as a practical description of what Meyer considers the difference between an AI demonstration and a production process. In his telling, generating a working video was not enough: the first complete version ran, but he judged its pacing and visuals inadequate. The response was to turn subjective criticism into specific measurements and then make those measurements part of the build process.
That approach also assigns separate responsibilities to the people and tools involved. Meyer says Claude generated the engine, Codex looked for problems, tests checked whether fixes addressed those problems, and a human made the decision to publish. These are rules from one creator’s workflow, not evidence that the same process will work for every AI production system. Still, readers building AI-assisted products can see how review and approval steps may catch problems that a successful run or a green test suite would miss.
Meyer reinforces that warning with examples from other experiments. He says a model in a legal-answer test completed 12 of 15 answers after a weight edit, compared with four before, while its legal score remained zero. He also says one review round had 12 of the author’s tests passing while four of a reviewer’s six checks failed. The examples underline his distinction between completion, test results and actual quality.
Products Worth Considering
DJI Mic 2 Wireless Microphone System (2 Transmitters, Receiver, Charging Case)
Apple MacBook Air 15-inch with M5 chip, 16GB memory, 512GB SSD, and Liquid Retina display (Midnight)
EcoFlow DELTA 3 Portable Power Station (1024Wh, 1800W)
As an affiliate, we earn on qualifying purchases.
How the Video Engine Developed
The video project moved through several stages in a short period. Meyer says he first translated a reference video’s general style into a written list, then had Claude build a separate drawing engine using colors from his own design system and an openly licensed handwriting font. Codex, from another company, reviewed the code. Meyer says the first review produced five findings, which were addressed with tests that failed on the old code before the fixes were applied.
After approving a short test, Meyer says he published it that evening. The next morning’s full video prompted a different assessment, leading to a measured redesign. The first video in the improved look, about the Nasdaq’s worst day in 19 months, went live a few hours after his critique, according to his account. The video accompanying the November 6 article uses the same production approach and has narration made with an AI voice clone of Meyer.
The article also describes an August monitoring incident. Meyer says a guard produced about 1,300 warnings in a log that nobody read, while an alert email did not go out. He uses that episode to argue that monitoring only helps when an alert reaches a person able to act. The source gives no independent audit or external evaluation of these incidents; the numbers and outcomes are presented as Meyer’s own build and incident notes.
“And honestly, it was slow and boring.”
— Thorsten Meyer, describing the first full video
Limits of the Reported Results
The account is a first-person report, not an independent assessment. Meyer says the measurements came from his build notes, incident notes or experiments published on his site, but the article does not provide an external review of the video metrics, the code-review findings or the legal benchmark. The reported frame-fill and camera-move figures cover only the first three scenes of the rebuilt video, not every scene or a broader comparison across projects.
The legal experiment is also described briefly: the source gives the completion counts and says the legal score remained at zero, but does not specify the benchmark’s scoring method, the model, or how the weight edit was performed. It is therefore unclear what the result says about legal performance beyond this test. The article does not report whether the revised production gates have prevented later failures or how often they stop builds.
Meyer says the first full video remains online unchanged, but the source does not independently verify the publication history or provide an outside comparison of the original and revised versions. The account describes one creator’s process; it does not establish that the tools or workflow will produce similar results elsewhere.
Whether the Gates Hold Up
The next test, based on Meyer’s account, is whether the new checks continue to catch weak scenes as more videos are made. His stated process gives the build a way to flag low frame fill and objects that do not match the narration before rendering, but the article does not say how often those checks will be reviewed or updated.
Readers can view the video linked with the November 6 article, which Meyer says uses the same method. Further evidence of the workflow’s effectiveness would include results across more projects, details about how the measurements are calculated, and follow-up on whether alerts reach a human when problems occur. No publication schedule or independent evaluation is provided in the source.
Key Questions
What did Thorsten Meyer publish?
He published a first-person account on November 6, 2026, describing how he developed a hand-drawn explainer video style and added review and quality checks after the first full video disappointed him.
What changed in the rebuilt video?
Meyer says the team measured frame fill and still time, then changed the engine to fill more of the frame, move the camera with the narration and keep drawn objects moving slightly. In the first three scenes, he reports frame fill rising from 15% to 66%, still time dropping from 29 seconds to zero, and camera moves rising from three to 27.
Which AI tools did he use?
Meyer says Claude built the drawing engine and OpenAI’s Codex reviewed the code. He describes Codex as a reviewer rather than a builder, and says a human still made the decision to publish.
Are the reported results independently verified?
The source presents them as Meyer’s own build notes, incident notes and experiments. It does not include an independent audit of the video measurements, code-review findings or legal benchmark.
What does the legal-model example show?
Meyer says a model completed 12 of 15 legal benchmark answers after a weight edit, compared with four before, while its legal score remained zero. The article does not provide enough information about the model or scoring method to draw broader conclusions about its legal ability.
Source: ThorstenMeyerAI.com
