🔍 Read the full analysis: From AI Prototype To Production: The Rules I Learned Along The Way on ThorstenMeyerAI.com
Turn your quiet moments into listening time
- Thousands of audiobooks, podcasts and originals
- Listen on your phone, tablet or Echo — also offline
- Cancel anytime
TL;DR
Thorsten Meyer says he built a hand-drawn video engine in one evening, then used code review, human approval and measured quality gates to improve it after its first full video fell short. A separate model-editing experiment improved task completion but left the legal score at zero, illustrating that passing or finishing tasks does not by itself establish quality.
Thorsten Meyer said he built a hand-drawn video engine in one evening and published a more polished video hours after revising its first full-length output, which he described as “slow and boring.” In a November 6 account on ThorstenMeyerAI.com, Meyer details the safeguards he used to move the tool from prototype to publication: a second AI code reviewer, a human approval step, measured quality checks and automated gates.
Meyer said the style idea came from a video he admired, but he set a clean-room rule to avoid copying it. The team recorded a short list of broad traits, including hand-drawn strokes, a continuous visual world and colored number accents. He said the engine used no drawings, assets, layouts or drawing code from the reference. Its colors came from his own design system, and the handwriting font was openly licensed.
According to Meyer, Claude built the engine and OpenAI’s Codex reviewed the code without contributing to the build. Codex raised five findings in its first review round; Meyer said each was fixed that evening and paired with a test that failed against the old code. He also watched a roughly one-minute test video and approved it for publication. The short version appeared in an article about local votes against data centers.
The first full video in the new style appeared the next morning. Meyer said he found it too empty, insufficiently connected to its narration and static after drawings finished. Rather than rely only on subjective tweaking, he said the team measured frame coverage, time spent still and camera movement. Across the first three scenes of the rebuilt video, he reported that filled-frame coverage rose from 15% to 66%, still time fell from 29 seconds to zero, and camera moves increased from 3 to 27. These are figures for those three scenes, not a broader performance benchmark.
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
Checks That Shape Published AI Video
Meyer’s account matters because it describes production readiness as a process beyond generating a working demo. His approach adds separate checks for code, visual quality and editorial judgment: another model reviews the implementation, tests are designed to catch specific defects, numerical thresholds block weak scenes, and a person decides whether the result should go live.
The distinction is practical for teams publishing AI-generated work. A system can run and still produce material that is dull, misleading or poorly matched to its narration. Meyer’s reported measurements also show how an aesthetic criticism can be translated into checks that a build process can enforce. They do not, by themselves, prove that viewers found the revised videos better; the account provides no audience study or comparative engagement data.
A separate experiment in the same article reinforces the limits of narrow success measures. Meyer said a model edited for a legal task completed 12 of 15 benchmark answers, compared with 4 before the edit, while its legal score remained zero. The higher completion count did not amount to a better legal result under the benchmark’s scoring.
Products Worth Considering
Bowflex SelectTech 840 Adjustable Kettlebell
Hyperice Venom 2 Heated Vibration Back Wrap
Blueair Blue Pure 211i Max Air Purifier for Large Rooms
As an affiliate, we earn on qualifying purchases.
From One-Minute Demo to Full Video
The account traces a short sequence of development. Meyer first translated inspiration from another creator’s video into a limited set of general visual traits, then had Claude build a new engine. Codex reviewed the code that evening. After Meyer approved a test clip, a roughly one-minute version went live. The first full video followed the next morning and prompted the redesign.
Meyer said the rebuilt approach made each shot fill more of the frame, tied camera movement to the narration, kept finished drawings subtly moving and required each object to illustrate something being said at that moment. He said the first video in the improved look was published hours after his critique, alongside coverage of the Nasdaq’s worst day in 19 months. The article also points to a separate published report about the legal-model experiment.
Meyer describes the work as a personal account, with numbers drawn from his build notes, incident notes and published experiments. The results therefore reflect his own projects rather than an independent evaluation of AI production methods.
“And honestly, it was slow and boring.”
— Thorsten Meyer, describing the first full video
What the Report Does Not Establish
The account does not provide independent verification of the video metrics, a defined scoring method for frame coverage or stillness, or evidence that the revised style improved audience response. The reported 15%-to-66% coverage change and other figures apply to the first three scenes only; results across the full video are not specified.
It is also unclear from the available material how the legal benchmark was constructed, what counted as a completed answer, or why the model’s legal score remained zero. Meyer says the August incident involved a guard that wrote about 1,300 warnings to a log nobody read and an alert email that did not go out, but the account does not give the guard’s purpose or say what corrective action followed. These examples are the author’s reports, not independently audited findings.
Quality Gates Before the Next Render
Meyer says the video engine now uses the measured results as gates: a scene judged too empty is stopped before rendering, and a build raises a warning if an object does not illustrate the narration at that moment. The article presents those checks as part of the current workflow, alongside code review, tests and human approval.
The account does not announce a product launch, release date or independent evaluation. The next steps visible in Meyer’s description are continued use of the gates on new videos and further reporting on the separate model-editing experiment. Whether the checks improve viewer response, and how the monitoring failure will be addressed, remain unanswered.
Key Questions
What did Thorsten Meyer build?
He said he built an engine for producing hand-drawn-style videos, using Claude to create it. He described a test clip and later full videos produced with the system.
How did the team improve the first full video?
After Meyer criticized its empty frames, weak connection to the narration and static images, the team measured frame coverage, still time and camera moves. He said it then redesigned the engine and made some quality thresholds automatic gates.
What role did Codex have?
Meyer said OpenAI’s Codex reviewed the code but did not build the engine. He reported five findings in the first review round and said each fix had a test that failed on the previous code.
Did the legal-model experiment improve the model?
Meyer said the edited model completed 12 of 15 legal benchmark answers, up from 4, but its legal score remained zero. The account does not provide enough benchmark detail to assess the result independently.
Are the reported video results independently verified?
No independent verification is provided in the account. Meyer says the figures come from his own build notes, incident notes or experiments, and the video measurements cited here cover the first three scenes.
Source: ThorstenMeyerAI.com
