🔍 Read the full analysis: Building AI For The Real World: Lessons From My First Video on ThorstenMeyerAI.com
Get business pricing on monitors, keyboards and dev gear
- Business-only prices and quantity discounts
- Tax-exempt purchasing
- Multiple users, one account, clear invoices
TL;DR
Thorsten Meyer says his new hand-drawn video engine moved from an evening prototype to a published video after he reviewed and measured its first full-length output. His account describes adding automated checks, a second AI code reviewer and human approval, while a separate legal-model experiment showed that completing more benchmark tasks did not improve the model’s legal score.
Thorsten Meyer has published an account of building a hand-drawn video engine and improving it after its first full video struck him as “slow and boring.” The November 6 post, updated November 9, says the revised workflow uses measured scene checks, a separate AI code reviewer and a human approval step before publication. The account matters as a practical case study in the gap between getting an AI-assisted system to work and deciding whether its output is good enough to release.
Meyer says the idea came from a hand-drawn explainer video he liked. He wrote down a limited set of stylistic traits—such as hand-drawn strokes, a continuous visual world and large numbers used as color accents—and says the new engine used no drawings, assets or layouts from the video that inspired it. Claude built the engine from scratch, according to Meyer, using his own design system and an openly licensed handwriting font.
For the first code review, Meyer says OpenAI’s Codex examined the work but did not build it. Five findings from that round were fixed, he writes, and each fix received a test designed to fail on the earlier code. Meyer then watched a test video of about a minute and approved it for publication. A later full video did not meet his expectations: he identified excessive white space, icons that did not support the narration and finished drawings that remained still. That version remains online, unchanged, as an example of the shortcomings.
Instead of relying only on subjective tweaking, Meyer says he measured frame coverage and time spent completely still, then made checks part of the build. Across the first three scenes of the rebuilt video, he reports that filled-frame coverage rose from 15% to 66%, still time fell from 29 seconds to zero, and camera moves increased from three to 27. These are figures from his own project notes, not an independently audited evaluation. He says the improved video, about the Nasdaq’s worst day in 19 months, was published a few hours after his critique.
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 Metrics to Release Gates
The account’s central point is that production quality requires more than a successful render. Meyer describes turning subjective feedback—too much empty space, unhelpful imagery and static scenes—into checks that can stop a scene from proceeding. That approach gives the team a repeatable way to catch some problems before a video is rendered, rather than relying on someone to notice them only after viewing the finished output.
His reported figures illustrate a change in the specific scenes he measured, but they do not establish how the system performs across a large set of videos. The practical value of the account is its emphasis on connecting each quality claim to a measurement and a review step. Meyer also stresses that the system’s automated checks do not replace a person’s decision to approve the result. For readers building AI-assisted products, that distinction matters: passing a test, finishing a task and meeting a human standard are different outcomes.
A second experiment in the post makes a related point about evaluation. Meyer says a local model answered more tasks after an AI edited its weights, but its legal score remained zero. More completed answers, by themselves, did not show better legal performance. That example reinforces the need to measure the result that matters, rather than treating activity or completion as a proxy for quality.
Products Worth Considering
Apple Watch SE 3 GPS 40mm Smartwatch
Bose TV Speaker (2nd Gen) Compact Soundbar
LG UltraGear 25G590B-B 25-Inch Full HD 1000Hz Gaming Monitor
As an affiliate, we earn on qualifying purchases.
How the Video Workflow Developed
Meyer frames the post as a personal account of his production process, with figures drawn from his build notes, incident notes or experiments published on his site. The initial prototype came together in one evening, he says. A one-minute test video was published inside an article about local votes against data centers after he watched and approved it. The next morning’s full video exposed problems that the shorter test had not resolved.
The November 9 update says the figures now refer to 10 published Lexassay tasks and that the runs are public on lexassay.com. Meyer’s post also describes an earlier monitoring failure: in August, a guard wrote about 1,300 warnings to a log nobody read, while an alert email did not go out. He uses that incident to argue that monitoring only helps when its warnings reach a person who can act on them.
The post says every real code-review finding should lead to a test that fails on the old code. Meyer gives another example of why he favors independent checks: in one review round, he writes, all 12 of his own tests passed, while four of a reviewer’s six checks failed. These examples are reported by Meyer; the post does not present them as independently verified results.
“A demo that runs is not a decision.”
— Thorsten Meyer
What the Report Does Not Establish
The video metrics cover only the first three scenes of the rebuilt video, according to Meyer. The post does not provide a larger evaluation showing whether the same results hold across many videos, nor does it give an independent assessment of the reported measurements. It also does not specify a full benchmark for judging whether viewers find the videos clearer or more engaging.
The legal-model result is similarly limited in the material provided: the model answered all 10 published tasks rather than four, but its legal score remained zero. The post does not establish why the score stayed at zero or whether the weight-edit approach would produce different outcomes on other models or tasks. Meyer’s account also does not identify a formal review standard for the human approval step. The reported workflow and results should therefore be read as one builder’s experience, not proof that the same methods will produce the same results elsewhere.
The Checks Before Future Videos
Meyer says the new measurements are now gates in the video build: a scene judged too empty can be stopped before rendering, and a check can flag an object that does not illustrate what is being said. The next step described in the post is continued use of that workflow in later videos, with a human still reviewing the output before it goes live.
The post does not announce a separate product launch, release date or broader performance study. Readers can compare the legal-model runs on lexassay.com, which Meyer says are public, and look to future published videos to see whether the scene checks remain part of the production process. Whether the reported improvements generalize beyond these examples remains open.
Key Questions
What did Thorsten Meyer build?
He says he built a hand-drawn video engine from scratch, using a set of visual traits as inspiration while avoiding the original video’s drawings, assets and layouts.
What changed after the first full video?
After describing the first full video as slow and boring, Meyer says he measured frame coverage and still time, revised the engine and added checks to stop scenes that are too empty or contain imagery unrelated to the narration.
Are the video improvement figures independently verified?
The figures come from Meyer’s own account and refer to the first three scenes of the rebuilt video. The post does not describe an independent audit or a larger evaluation across multiple videos.
Did the legal-model experiment improve the model’s legal score?
No, not according to Meyer. He says the model answered all 10 published tasks instead of four, while its legal score stayed at zero.
What role did human review play?
Meyer says he watched a roughly one-minute test video and approved it before publication. He describes human approval as a required decision, separate from a demo running or automated checks passing.
Source: ThorstenMeyerAI.com
