🔍 Read the full analysis: The Small, Boring First Step That Took My AI Idea 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 one evening, then published a short test after a separate AI reviewed the code and he approved the result. The first full video fell short, prompting measured changes and checks that now block some flawed scenes; the account is a personal report, not an independently audited evaluation.
Thorsten Meyer says he took a new hand-drawn AI video style from an evening prototype to a live site within hours, using separate AI code review, his own approval and later automated checks to address shortcomings in the first full video. In a personal account published November 6, he describes the process as a case study in moving an AI idea into production, while the results and measurements come from his own build notes and experiments.
Meyer says he drew inspiration from another video’s style but set a clean-room rule: his team wrote down broad traits, such as hand-drawn strokes and a continuous visual world, without using the original’s drawings, assets or layouts. He says Claude built a new engine from scratch, using his own design system and an openly licensed handwriting font.
In his workflow, Claude built the code and OpenAI’s Codex reviewed it without contributing to the build. Meyer reports that Codex found five issues in its first review. He says those findings were fixed that evening and that each fix was paired with a test that failed against the earlier code.
Meyer watched a roughly one-minute test video and approved it for publication. The following morning, however, the first full video felt slow and unengaging to him. He identified excessive white space, icons that did not match the narration and drawings that stopped moving once complete. He says that original video remains online as an example of the problems.
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 Checks
The account’s central point is that publication was not treated as proof of quality. Meyer says he translated his criticism into measurements: how much of the frame contained content, how long an image stayed still and how often the camera moved. For the first three scenes of the rebuilt video, he reports that filled-frame coverage rose from 15% to 66%, still time dropped from 29 seconds to zero, and camera moves increased from three to 27.
Those figures describe three scenes in Meyer’s own project, not an independently tested benchmark or a general measure of AI video quality. He says the measurements now function as gates: scenes judged too empty are stopped before rendering, while the build flags objects that do not illustrate the narration. That turns a subjective critique into a repeatable production check, although it cannot by itself establish that a video is clear, accurate or useful to viewers.
The workflow also separates responsibilities: one model generates code, another reviews it, tests check reported findings, and a person responsible for the output makes the release decision. For readers building AI tools, the example shows one practical approach to reducing the gap between a working demo and a reviewed product. It does not establish that the same checks will work for other systems or use cases.
Products Worth Considering
Samsung Galaxy Tab A11 8.7-Inch Wi-Fi Tablet, 64GB, 4GB RAM, Silver
Bose QuietComfort True Wireless Earbuds with Active Noise Cancellation, Black
EcoFlow DELTA 3 Portable Power Station (1024Wh, 1800W)
As an affiliate, we earn on qualifying purchases.
A Prototype, Then a Full Video
Meyer places the initial idea and the first short test on the same evening. The test appeared within an article about local votes against data centers. The first complete video followed the next morning, and his critique prompted a rebuild. He says a video in the revised style about the Nasdaq’s worst day in 19 months went live a few hours after that critique; the source does not provide a publication date for that market event.
The account also describes a separate experiment from the same week. Meyer says Claude edited the weights of a local model for a legal task. In his benchmark, the model completed 12 of 15 answers, compared with four before the edit, while its legal score remained zero. The source presents this as a reminder that finishing more answers does not necessarily mean a better result. It supplies no further benchmark details in the provided material.
Meyer connects the process to an earlier monitoring failure: he says a guard generated about 1,300 warnings in August, but nobody read the log and an alert email did not go out. That episode reinforces a point in his account: monitoring only helps when someone receives and acts on it.
“I really like this. It can go live.”
— Thorsten Meyer, describing his reaction to the test video
What the Account Does Not Establish
The source is Meyer’s personal account; the supplied material does not include independent verification of the code review, the video measurements or the legal benchmark. It does not give the full scoring method, the benchmark’s contents, the legal score’s scale or details needed to reproduce either experiment.
The reported video changes cover the first three scenes, and the account does not say how many scenes were tested overall or how the thresholds for stopping a scene were set. It also does not report audience response, viewing data or whether the gates prevent problems beyond the specific issues Meyer identified. The monitoring example does not specify what the warnings concerned or when the missed email was expected to arrive.
The Gates Will Face More Builds
Meyer says the new checks are now part of the video build process, but the account does not announce a formal release schedule or a next evaluation date. The immediate next test is whether the frame-fill, movement and narration-match checks continue to flag problems as the engine produces more scenes and videos.
Readers can compare the first full video, which Meyer says remains online unchanged, with later work in the revised look. Further evidence would include results across a larger set of videos, clearer benchmark methods and information about whether the checks catch issues that human reviewers would otherwise miss.
Key Questions
What did Thorsten Meyer put into production?
He describes a hand-drawn video engine that turns narration into a continuous illustrated scene, with camera movement and animated drawings. He says a short test was published first, followed by a revised full video after he identified problems in the initial version.
What roles did Claude and OpenAI’s Codex have?
According to Meyer, Claude built the engine, while OpenAI’s Codex reviewed the code and reported findings. Meyer says Codex did not build the system and that fixes for its first five findings each received a test that failed against the earlier code.
What changed after the first full video?
Meyer says he measured frame coverage, still time and camera movement, then changed the engine to fill more of the frame, move with the narration and keep completed drawings slightly animated. For the first three scenes, he reports coverage rose from 15% to 66%, still time fell from 29 seconds to zero, and camera moves rose from three to 27.
Are the reported results independently verified?
The supplied source is a first-person account based on Meyer’s own notes and experiments. It does not provide independent verification or enough benchmark detail to reproduce the reported results.
Source: ThorstenMeyerAI.com
