TL;DR
Get business pricing on office and shipping supplies
- Business-only prices and quantity discounts
- Tax-exempt purchasing
- Multiple users, one account, clear invoices
Thorsten Meyer says he built a hand-drawn video engine in one evening after documenting style traits from a video he liked. The first full video was slow and visually sparse; after measuring and revising the opening scenes, he published an improved version and added automated gates. His account also describes limits in AI code review, model scoring and monitoring.
Thorsten Meyer says he built a hand-drawn video engine in one evening and published an improved video hours after its first full version disappointed him, describing a workflow that used separate AI tools, human approval and measured checks. His account, published November 6, 2026, offers a personal case study in moving an AI-assisted creative tool from prototype to production, while its results remain specific to his own project.
Meyer says the idea came after watching a video whose visual style he liked. To avoid copying it, he says his team used a clean-room approach: they recorded a short list of style traits, such as hand-drawn strokes and a continuous visual world, without using the source video’s drawings, assets or layouts. Claude built the engine from scratch, according to Meyer; the colors came from his design system and the handwriting font was openly licensed.
He says OpenAI’s Codex reviewed the code but did not build it. Codex identified five findings in the first review round, and Meyer says each fix came with a test that failed against the old code. Later that evening, he watched a roughly one-minute test video and approved it for publication in an article about local votes against data centers.
The first full video in the new style was ready the next morning, but Meyer judged it slow and visually dull. He then measured the opening three scenes and revised them. His build notes report that the content-filled share of the frame rose from 15% to 66%, time spent with a still picture fell from 29 seconds to zero, and camera moves increased from three to 27. Those figures describe the first three scenes in his project, not a general benchmark.
From AI idea to production in one evening
Thorsten Meyer turned a hand-drawn video idea into a working engine, then used human review and measurable checks to improve the result before publication.
Build with clear roles
Separate tool jobs and keep the publication decision with a person.
Claude made the engine
The engine was built from scratch using style traits, without copying the source video’s drawings, assets or layouts.
Codex checked the code
Codex reviewed the code; Meyer says it identified five findings in the first round.
A person watched the output
Meyer approved a short test video, then reviewed and revised the first full video before publication.
Turn critique into gates
The first full video looked slow and visually dull, so Meyer measured its opening scenes and changed them.
Identify concrete problems in the scenes viewers will see.
Track filled frame area, still-picture time and camera moves.
Stop later builds when scenes are too empty or icons do not illustrate the narration.
What the checks can miss
Passing one measure does not prove the whole system is reliable.
- Tests can show a fix catches a specific old bug.
- Scene gates make repeated quality checks possible.
- Human review evaluates the published result.
- Author-written tests passed while some reviewer checks failed.
- A model improved its legal-answer count while its legal score stayed zero.
- About 1,300 monitoring warnings went unread, and an alert email did not go out.
From Prototype to Published Video
The account shows that getting a working AI tool to produce an output did not settle whether that output was ready to publish. Meyer says a human watched and approved the short test video, then reviewed the first full video and identified problems that could be translated into measurable changes. The distinction matters for teams adopting AI in creative work: completion is not the same as quality, and approval needs to cover the result people will actually see.
Meyer says two thresholds now stop later builds when scenes are too empty or icons do not illustrate what is being said. That turns subjective feedback into checks his process can repeat. The measurements are narrow and self-reported, but they show how a critique can become a production rule rather than remain an informal impression.
His other examples point to limits in treating automated checks as proof. In one code review round, he says all 12 tests written by the author passed, while four of six checks from the reviewer failed. In a separate weight-edit experiment, a model completed 12 of 15 legal answers instead of four, but its legal score remained zero. These examples are also from Meyer’s notes; they underscore that passing tests or improving one count may not establish that a system is reliable overall.
Products Worth Considering
Apple MacBook Neo 13-inch Laptop with A18 Pro Chip, 8GB Memory, 256GB SSD, Liquid Retina Display, and 1080p FaceTime HD Camera – Blush
HP OfficeJet 250 Wireless Mobile All-in-One Inkjet Printer
Apple MacBook Pro MPHF3LL/A 14.2-Inch Laptop | M2 Pro 12-Core Chip 16GB Unified Memory 1TB SSD | 19-Core GPU Liquid Retina XDR Display Space
As an affiliate, we earn on qualifying purchases.
The Checks Behind the Build
Meyer presents the video project as part of a broader practice for taking AI ideas into production, and says the figures in the post come from his own build notes, incident notes or published experiments. The video timeline is given in relative terms: the idea and initial engine came in one evening, the first full video was ready the following morning, and the revised version went live hours later. The post does not give a calendar date for each step beyond saying the work happened that week.
His process assigns different roles to the tools and people involved. Claude builds the engine, Codex reviews code, tests are meant to expose specific bugs, and Meyer makes the publication decision after watching the output. He argues that tests should be shown to fail on old code before a fix is accepted, so the test demonstrates that it can detect the original problem.
He also cites an August monitoring incident: a guard produced about 1,300 warnings in a log that nobody read, and an alert email did not go out. The post does not specify the duration covered by those warnings. Meyer uses the incident to argue that monitoring only helps if it reaches a person who can act on it.
“A demo that runs is not a decision.”
— Thorsten Meyer
What the Build Notes Do Not Show
The account is a first-person report based on Meyer’s own project records. It does not provide independent verification of the measurements, a comparison with other video engines, or enough detail to assess how the two scene gates perform across a larger set of videos. The post also does not state how much time or compute the work required, or whether the workflow has been repeated successfully.
The review and test examples raise questions about reliability, but the post does not identify the nature or severity of the five code findings, the four failed reviewer checks, or the model’s 15 legal-answer prompts. It also does not explain what the model’s legal score measures. The August warning incident is described without a full account of the underlying guard or alert system. These details limit what readers can conclude beyond Meyer’s own experience.
How the Production Gates Apply
Meyer says the video process now includes two gates: one for scenes that are too empty and another for icons that do not illustrate the narration. The post does not announce a future release date or a broader rollout. The next practical test is whether those checks continue to catch the same problems in subsequent videos and whether a human review remains part of the publication decision.
For readers following the project, the post’s next evidence would be additional build notes showing how the gates perform across later videos, along with details on any changes to the engine or review process. Meyer has not yet provided that information in this account.
Key Questions
What did Thorsten Meyer build?
He says he built an engine for producing hand-drawn-style videos, using Claude to write the engine and Codex to review its code.
How did the first full video change?
For its first three scenes, Meyer reports that filled frame area rose from 15% to 66%, still-picture time fell from 29 seconds to zero, and camera moves rose from three to 27.
Was the video published without human review?
No. Meyer says he watched and approved a roughly one-minute test video before it went live. He later critiqued the first full video and revised it.
Do the measurements prove the workflow works for other teams?
No. The figures are from Meyer’s own project notes and cover the first three scenes in one video. The post does not provide independent verification or results across other teams or projects.
What remains unclear about the process?
The post does not give independent validation, detailed test prompts, the severity of review findings, or evidence about how the production gates perform across later videos.
Source: Thorsten Meyer AI
