🔍 Read the full analysis: What My First Evening Building With AI Taught Me About 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 one evening, but its first full video was slow and visually dull. He describes how critique, measured changes, code review and automated checks improved the output—and why passing tests or completing tasks does not by itself prove quality.
Thorsten Meyer says he built a hand-drawn video engine in one evening, then found its first full video slow and visually dull the next morning. In an account published on ThorstenMeyerAI.com and updated November 9, he describes how human approval, a second AI code review and measurable quality gates changed the production process—an example of why a working AI demo is not the same as a reliable result.
Meyer says the idea began after he watched a hand-drawn explainer whose visual style appealed to him. He set a clean-room rule for his own project: record a short list of broad style traits, but use no drawings, layouts, assets or drawing code from the video. The new engine used his own design system and an openly licensed handwriting font. He says Claude built it, while OpenAI’s Codex reviewed the code without contributing to the build.
In the first review round, Meyer says Codex reported five findings that were fixed that evening. He says each finding received a test written to fail on the earlier code, rather than a test that merely passed after the change. Later, he watched a roughly one-minute test video and approved it for publication. The following morning, however, the first full video in the new style disappointed him: it had too much empty space, icons that did not support the narration and drawings that stopped moving when complete. Meyer says that version remains online as an example of what went wrong.
Rather than revise the video by feel alone, Meyer says his team measured frame coverage, still time and camera movement, then built checks around the results. Across the first three scenes of the rebuilt video, he reports that frame coverage rose from 15% to 66%, still time fell from 29 seconds to zero, and camera moves rose from 3 to 27. He says the resulting rules now stop scenes that are too empty and flag objects that do not illustrate the spoken content. The figures are his own measurements, not an independently verified 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 Video Demo to Quality Gates
The account illustrates a production problem that applies beyond video: task completion is not a quality measure. An engine can render a file, and code can pass its tests, while the result still fails a human viewer or misses the intended purpose. Meyer’s response was to turn specific criticism—empty frames, static images and mismatched icons—into checks that can catch similar shortcomings earlier.
His description also assigns distinct roles to the tools and the person responsible for publication. Claude generated the engine; Codex reviewed code; tests checked reported issues; and Meyer made the release decision. That separation matters because an AI review is not the same as human approval, and automated checks only address what they have been designed to detect. Meyer’s experience is a first-person account, not a controlled comparison showing that the process will produce the same results in other projects.
Products Worth Considering
Amazon Kindle Paperwhite (2024, 16 GB) – 7-inch Glare-Free Display, Black
Logitech MX Mechanical Wireless Illuminated Keyboard, Tactile Quiet, Graphite
DJI Mini 4 Pro Fly More Combo Drone with DJI RC 2
As an affiliate, we earn on qualifying purchases.
The First Video Missed the Mark
The timeline in Meyer’s account spans an evening and the following morning. He says the initial idea became a working engine in one evening, and a short test video was approved and published that night within an article about local votes against data centers. The first full video using the new look arrived the next morning. Its weaknesses prompted a revised build and a second video, which Meyer says went live a few hours after his critique.
Meyer connects the video work to a separate experiment from the same week. He says a model modified through weight editing answered all 10 published legal benchmark tasks, up from four, while its legal score remained zero. He presents that result as a warning about treating a higher completion count as proof of improved performance. The source says the experiment runs are public on lexassay.com, but the account does not provide enough benchmark detail here to assess the scoring method independently.
“A demo that runs is not a decision.”
— Thorsten Meyer
What the Report Does Not Establish
The measurements come from Meyer’s own build and incident notes. The account does not describe an independent review of the video metrics, specify how the frame-coverage calculation was made, or show whether the new gates consistently improve later videos. The reported movement from 15% to 66% applies to the first three scenes of one rebuilt video, not to a broader sample.
It is also unclear from the account how the legal benchmark’s zero score is defined, why the model’s task completion increased, or whether the result has been independently reproduced. The claim that no material was copied from the original video is Meyer’s description of his process; the account does not report an external assessment of that question. These limits make the piece useful as a description of one workflow, not proof that its approach guarantees production quality.
Checks for Future Video Builds
Meyer says the revised rules are now part of the video build: checks can stop scenes judged too empty and flag objects that do not match the narration. The next practical test is whether those gates continue to catch problems as new videos are produced, and whether human review still finds issues that automated measures miss.
The legal-model experiment has a separate public record: Meyer says the runs for the 10 published tasks are available on lexassay.com. His account does not name a later review date or announce a broader evaluation. For readers, the immediate takeaway is to treat the reported numbers as project-specific evidence and watch for further results that show how the checks perform beyond these examples.
Key Questions
What did Thorsten Meyer build?
He says he built a hand-drawn video engine in one evening, using his own visual system rather than copying the source video’s drawings, layouts or assets.
What changed after the first full video?
After finding excessive empty space, static drawings and unhelpful icons, Meyer says he measured frame coverage, still time and camera movement, then added checks intended to catch similar issues before rendering.
How did the reported video measurements change?
For the first three scenes of the rebuilt video, Meyer reports frame coverage increased from 15% to 66%, still time dropped from 29 seconds to zero, and camera moves rose from 3 to 27. These are his project measurements, not an independently verified benchmark.
Did the legal-model experiment show that the model improved?
Meyer says the edited model answered 10 published legal tasks rather than four, but its legal score stayed at zero. The task count alone does not establish improved legal performance.
What remains unverified?
The account does not provide independent validation of the video measurements, evidence that the quality gates work across a larger set of videos, or enough detail to assess the legal benchmark and its scoring independently.
Source: ThorstenMeyerAI.com
