🔍 Read the full analysis: How I Took An AI Idea To Production In One Evening on ThorstenMeyerAI.com
Get hardware and tech essentials 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 explainer-video engine in one evening and published a short test after reviewing it himself and using OpenAI Codex to check Claude’s code. The first full video was slow and visually sparse, prompting measurable changes and automated checks; the account is a personal report, not an independent audit.
Thorsten Meyer says he built a new hand-drawn video engine in one evening, then published a short test after a second AI reviewed the code and he approved the result. In an account published November 6, he also describes how the first full video failed to meet his expectations, leading to measured design changes and automated checks before later videos were rendered.
Meyer, writing on ThorstenMeyerAI.com, says he used Claude to build the engine and OpenAI’s Codex as a separate reviewer. He says Codex did not write the system; it inspected the code and reported potential problems. Five findings from the initial review were fixed that evening, he writes, with a test for each finding that failed against the older code.
The initial design drew on broad style traits Meyer liked in another video: hand-drawn lines, a continuous visual world and large numbers used as colour accents. He says the team avoided taking drawings, assets, layouts or code from that work. The engine used his own design system, including blue for imagery and orange to highlight a key point, along with an openly licensed handwriting font.
Meyer says he watched a roughly one-minute test video and approved it for publication. It went live that evening within an article about local votes against data centers. The following morning, however, the first full video in the new style struck him as “slow and boring,” with excessive white space, icons that did not support the narration and images that stopped moving after they were drawn.
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 practical point is that a working demo and a production-ready result are different milestones. Meyer describes a process in which human approval came before publication, then concrete shortcomings in the longer video became targets for changes and checks. That distinction matters to teams using generative tools: a system can run and still produce work that misses its intended purpose.
After measuring the opening three scenes of the rebuilt video, Meyer reports that the share of the frame occupied by content rose from 15% to 66%, time spent completely still fell from 29 seconds to zero, and camera moves increased from three to 27. He says those measurements were made into build gates: a scene that is too empty is stopped before rendering, while an object that does not illustrate something in the narration triggers a warning.
The figures are the author’s own measurements, and the account does not provide an independent evaluation or a detailed technical specification. Still, the described approach offers a concrete example of converting subjective feedback into checks that can be repeated. Meyer’s wider point is that code review, human sign-off and monitoring only help when findings lead to tests or reach someone who can act on them.
Products Worth Considering
Logitech MX Anywhere 3S for Mac Compact Wireless Mouse, Space Gray
Apple MacBook Pro 14-inch (2021) – M1 Pro Chip, 10-Core CPU, 16-Core GPU, 16GB RAM, 1TB SSD, Space Gray
Logitech MX Master 3S Bluetooth Wireless Mouse, Graphite
As an affiliate, we earn on qualifying purchases.
The First Video Was the Test
Meyer says the first one-minute version was published the evening the engine was built. The longer video made the next morning exposed problems the short test had not resolved. Rather than quietly replace that version, he says it remains online as an example of what did not work. A later video, on what he describes as the Nasdaq’s worst day in 19 months, used the revised look and went live a few hours after his critique.
He connects the production account to two other examples from his work. In a separate experiment involving edits to a model’s weights, a model completed 12 of 15 legal benchmark answers, compared with four before the edit, while its legal score remained zero. Meyer presents that outcome as a warning that finishing more answers does not establish correctness. He also recalls an August monitoring failure: a guard produced about 1,300 warnings in a log nobody read, and an alert email did not go out.
These are examples from Meyer’s own projects, not independently verified findings. The video accompanying the account is described as a roughly four-and-a-half-minute hand-drawn explainer with a voice clone of Meyer. The source does not identify the specific benchmark, explain how its score is calculated or provide a full account of the model-weight experiment.
““I really like this. It can go live.””
— Thorsten Meyer
What the Account Does Not Establish
The account is a first-person description of Meyer’s workflow and measurements. It does not include an independent review of the code, the video-quality figures or the reported benchmark results. The source also does not specify the exact thresholds used by the new scene gates, the full evaluation method for the legal benchmark or whether the checks were tested across a larger set of videos.
It is also unclear how much of the production process was automated beyond generating and checking scenes, or how the engine performs on subjects and formats unlike the examples described. Meyer reports that the first test and later videos were published, but gives no audience, accessibility or production-cost data. Those omissions limit what can be concluded about the system’s broader reliability or impact.
How the New Gates Will Be Used
Meyer says the frame-fill and motion measurements now act as checks before rendering, while the system flags objects that do not correspond to the narration. The account does not announce a future release date, external audit or formal evaluation. The next evidence of how well the approach works will come from later videos and whether the gates catch problems without introducing new ones.
For now, the published account records a short path from idea to a live test, followed by a failed first full version and a revised production process. Whether that process generalizes beyond Meyer’s own site remains unestablished.
Key Questions
What did Thorsten Meyer build?
He says he built a hand-drawn explainer-video engine in one evening, then used it to publish a short test video and later a revised full video.
Which AI tools did he use?
Meyer says Claude built the engine, while OpenAI’s Codex reviewed the code and reported findings. He describes Codex as a reviewer rather than a builder.
What changed after the first full video?
After finding too much white space and static imagery, Meyer says he measured frame coverage, still time and camera movement. For the first three scenes of the rebuild, he reports coverage rising from 15% to 66%, still time falling from 29 seconds to zero, and camera moves rising from three to 27.
Are the results independently verified?
The source is Meyer’s personal account and does not provide independent verification of the code review, video measurements or other experiments.
Source: ThorstenMeyerAI.com
