🔍 Read the full analysis: The Reality Of Taking An AI Idea From Sketch To Production 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 an evening, but the first full video was slow and visually dull. After measuring scene composition and motion, he added automated gates and human review; his account also describes a separate AI experiment whose improved completion rate did not improve its legal score.
Thorsten Meyer says he turned an idea for a hand-drawn AI video style into a working engine in one evening, then revised it after the first full video proved slow and visually dull. In an account published November 6 on ThorstenMeyerAI.com, he describes the changes that made the format usable on his site: code review by a second AI, human approval, measured production checks and automated gates.
Meyer says the idea came from a video whose visual style he admired. To avoid copying it, he wrote down a short set of broad traits—hand-drawn strokes, a continuous visual world and large numbers used as colour accents—and says he used no drawings, assets or layouts from the original. Claude built the engine from scratch, according to Meyer; the colours came from his design system, and the handwriting font was openly licensed.
He assigned OpenAI’s Codex a separate role: reviewing rather than building. Meyer says Codex identified five issues in its first review, which were fixed that evening. He says each finding was paired with a test that failed on the old code, a method intended to show that the test could detect the problem rather than merely pass after a fix.
A roughly one-minute test video received Meyer’s approval and was published that evening within an article about local votes against data centers. The next morning, however, the first full video in the new style felt “slow and boring” to him. He identified excess white space, icons that did not support the narration and completed drawings that remained motionless. Rather than replace that version, he says it remains online 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
How the Prototype Became a Workflow
The account illustrates a practical distinction for teams using generative AI: a working demo is not the same as a production-ready result. Meyer’s description puts the work after the initial build—review, evaluation, approval and repeatable checks—at the centre of moving an AI-assisted idea into regular use.
His example also shows how a subjective critique can be turned into concrete production criteria. After identifying the visual problems, Meyer says he measured frame coverage and still-image duration, then changed the engine so that scenes filled more of the frame, camera movement followed the narration and objects continued to move subtly after being drawn. He says the same checks now stop scenes that are too empty and flag visuals that do not illustrate the spoken content.
Those controls matter because a polished-looking output can still be misleading or irrelevant, and passing software tests alone does not establish that the result is useful. Meyer’s account argues for human judgment alongside automated checks, rather than treating a model’s completion or a green test suite as a final sign-off.
Products Worth Considering
LG UltraGear 34GX90SA-W 34-inch WQHD Curved OLED Gaming Monitor
LG UltraGear 32GX850A-B 32-inch 4K UHD Glossy OLED Gaming Monitor
LG UltraGear 32GS60QC-B 32-inch Curved QHD Gaming Monitor (180Hz, 1ms, 1000R, AMD FreeSync, HDR10)
As an affiliate, we earn on qualifying purchases.
From First Cut to Measured Revision
On the first three scenes of the rebuilt video, Meyer reports that the share of the frame filled with content rose from 15% to 66%. He says time with a completely still image fell from 29 seconds to zero, while camera moves increased from three to 27. These figures apply to those three scenes, according to his account; they are not a general benchmark for AI video production.
Meyer says the improved-look video went live a few hours after his critique, accompanying an article about the Nasdaq’s worst day in 19 months. The video embedded with his account uses the same process and runs about four and a half minutes, with narration by an AI voice clone of Meyer.
He also cites a separate experiment involving a local model and legal answers. After Claude edited the model’s weights, it completed 12 of 15 answers, compared with four before, but its legal score remained zero. Meyer presents this as a warning that finishing more answers does not necessarily mean improved accuracy or quality.
“A demo that runs is not a decision.”
— Thorsten Meyer
Limits of the Reported Results
The measurements and outcomes are Meyer’s own reported results, drawn from his build notes, incident notes and experiments, rather than an independently evaluated study. The account does not provide enough detail to assess the measurement method, the full benchmark design or how performance varies across other scenes and tasks.
It is also unclear how often the new gates reject scenes, how frequently human reviewers override them, or whether the revised workflow has been tested at a larger production scale. The reported legal benchmark result is limited to 15 answers; the source does not establish why the score stayed at zero or whether the weight edit affected other measures.
Meyer’s account mentions an August monitoring failure in which a guard reportedly wrote about 1,300 warnings to an unread log and an alert email did not go out. It does not give the full circumstances or say whether the monitoring system has since been changed. The figure is a raw count; no time window or comparison baseline is specified.
Checks Before the Next Release
Meyer says the measured video criteria are now built into the production process: a scene that is too empty can be stopped before rendering, while an object that does not match the narration triggers a warning. The account does not set out a dated release plan or name a next milestone beyond using the workflow on subsequent videos.
For readers following the project, the next useful evidence would be whether the gates continue to catch problems in later productions and whether the alerts reach a person who can act on them. Further details about the legal-model experiment would also be needed to interpret its unchanged score. For now, the published account offers a single creator’s process and measurements, not proof that the same setup will produce similar results elsewhere.
Key Questions
What did Meyer build?
He says he built an engine for hand-drawn explainer videos, with a continuous visual world, camera movement and illustrated narration, then used it to make videos for his site.
Why did the first full video get revised?
Meyer said it had too much white space, unhelpful icons and long stretches without movement. He measured frame coverage and still-image duration, then revised the engine and made those criteria part of its production checks.
What did the reported video measurements show?
For the first three rebuilt scenes, Meyer reported frame coverage rising from 15% to 66%, still-image time falling from 29 seconds to zero, and camera moves increasing from three to 27. Those figures describe his scenes, not a broader industry benchmark.
Did the legal-model experiment improve the model?
It completed more benchmark answers after a weight edit—12 of 15 rather than four, according to Meyer—but its reported legal score remained zero. The account does not establish broader improvement.
Are the results independently verified?
The article describes Meyer’s own build notes and experiments. It does not report an independent audit or provide enough methodological detail to verify the results from the account alone.
Source: ThorstenMeyerAI.com
