🔍 Read the full analysis: AI Production Lessons From My First, Very Ordinary Video on ThorstenMeyerAI.com
Get hardware and tech essentials 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 his first full video from a new AI drawing engine was slow and visually static, despite an earlier short demo he liked. He revised the engine using measured visual checks and added gates to block scenes that miss them; a separate legal-model experiment showed why completion counts alone can mislead.
Thorsten Meyer says a new AI drawing engine produced a slow, boring first full video the morning after he built it, prompting him to measure visual problems and add automated checks before rendering. In a November 6 account updated November 9, Meyer describes a workflow combining AI code review, human approval and production gates—and argues that a system completing a task is not proof that its result is good.
Meyer says he built the engine in one evening after seeing a hand-drawn explainer style he liked. He describes setting a clean-room rule: record broad traits such as hand-drawn strokes and a continuous visual world, but do not reuse the original video’s drawings, assets, layouts or code. The new engine used his own design system, with blue imagery, orange accents and an openly licensed handwriting font.
According to Meyer, Claude built the engine, while OpenAI’s Codex reviewed the code without contributing to the build. Codex identified five issues in its first review round. Meyer says each fix was paired with a test that failed against the old code, so the test could show that it detected the original problem. He also says a roughly one-minute test video went live after he watched and approved it.
The next morning’s full video did not meet his expectations: he cited excessive white space, icons that did not support the narration and images that stopped moving after they were drawn. Meyer measured frame coverage and still-image duration, then changed the engine. Across the first three scenes, he reports that filled-frame coverage rose from 15% to 66%, time standing still fell from 29 seconds to zero, and camera moves increased from 3 to 27. These figures refer to those scenes, not necessarily the entire video. The revised checks now block scenes he considers too empty and flag objects that do not illustrate the spoken content.
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 Visual Critique to Build Gates
The account offers a practical example of the gap between an AI demo and a production process. Meyer’s first short test prompted a human approval, but the full video still exposed problems that were not captured by the initial reaction. His response was to turn subjective criticism into measurable checks and make those checks part of the build, rather than relying on repeated adjustments by taste alone.
The distinction matters for teams using generative tools: a video can render successfully without being clear or engaging, just as a model can produce answers without performing well on the measure that matters. Meyer’s numbers describe his own project and should not be treated as independent evidence that the approach will improve other systems. The broader lesson supported by his examples is narrower: completion, passing tests and quality are different outcomes, and each needs an appropriate review.
His account also places people inside the workflow. A second AI can inspect code, and automated rules can stop a build, but Meyer says a responsible person still has to watch the result and approve publication. That human decision is part of the production process, not a substitute for measurement or a guarantee of quality.
Products Worth Considering
Ledger Flex + Folio Crypto Hardware Wallet
Ledger Nano X Bluetooth Crypto Wallet
Apple MacBook Air 13-inch Laptop with M5 Chip, 16GB Memory, 512GB SSD, Liquid Retina Display, and 12MP Center Stage Camera (2026, Silver)
As an affiliate, we earn on qualifying purchases.
Two Projects, One Testing Lesson
Meyer presents the video work alongside a separate experiment involving a local model and a published legal benchmark. He says Claude edited the model’s weights and that the model then answered all 10 published tasks, up from four. However, its reported legal score remained zero. The update says the figures refer to the 10 published Lexassay tasks and that the runs are public on lexassay.com.
The legal experiment is not evidence about the video engine, and the article does not provide enough detail to assess the benchmark or the weight-edit method independently. Meyer uses it to illustrate the risk of selecting a convenient metric: more completed answers did not mean a better legal score in his test. He also reports that, in one code-review round, all 12 of his own tests passed while four of a reviewer’s six checks failed. In another example from August, a guard reportedly logged about 1,300 warnings without reaching a person because its alert email did not go out.
““A demo that runs is not a decision.””
— Thorsten Meyer
What the Report Does Not Establish
The account is a first-person report based on Meyer’s own build notes, incident notes and experiments; it is not an independent evaluation of the workflow. The reported visual metrics cover the first three scenes, and the article does not specify the exact measurement method, how the thresholds were selected or whether the improvements were assessed by viewers beyond Meyer’s own review.
The article also does not provide a controlled comparison showing how the revised engine performs across other videos, topics or production teams. Details of the legal-model experiment—including its method and the meaning of a score of zero—are not fully explained in the supplied account. The reported warning and test figures likewise describe specific incidents, not rates across a broader period.
Checks Before Future Videos
Meyer says the visual measurements have been turned into build-time gates: scenes that are too empty should be stopped before rendering, while objects that do not match the narration should trigger a warning. Future work will depend on whether those checks catch problems consistently and whether the finished videos meet human review, rather than only passing automated rules.
The source says the improved-look video about the Nasdaq’s worst day in 19 months went live a few hours after Meyer’s critique, but it gives no publication date for that event in the supplied material. Meyer also says a further video uses the same process. No additional results or independent assessment are provided, so the account leaves open how the workflow performs over time.
Key Questions
What happened with Meyer’s first full AI-generated video?
Meyer says it came out the morning after he built the drawing engine and felt slow and boring to him, with excessive white space, weakly relevant icons and static finished drawings. He says that version remains online unchanged as an example of what he wanted to improve.
What changed in the revised video engine?
Meyer says he rebuilt it to fill more of the frame, move the camera with each sentence, keep drawings in motion and connect each object to what is being said. He reports changes across the first three scenes, including coverage rising from 15% to 66% and still time falling from 29 seconds to zero.
Who reviewed the code and who approved the video?
According to Meyer, Claude built the engine and OpenAI’s Codex reviewed the code. Meyer says he watched the test video and made the publication decision himself; the article does not describe an outside approval process.
Does the legal-model experiment show that the model improved?
Not on the score Meyer reports. The model answered more of the 10 published tasks after a weight edit, but its legal score stayed at zero. The account does not provide enough methodological detail for an independent assessment.
Are the reported video measurements independent results?
No independent assessment is described in the source. The figures are Meyer’s reported measurements for the first three scenes, and the article does not specify the measurement method or show whether the changes generalize to other videos.
Source: ThorstenMeyerAI.com
