🔍 Read the full analysis: AI Production Lessons From A First Video Built In One Evening on ThorstenMeyerAI.com
TL;DR
Thorsten Meyer says he built a hand-drawn video engine in one evening, but its first full video was slow and visually static. After measuring the problems, he changed the production rules and made those checks gates; the figures and account come from his own build notes and experiments.
Thorsten Meyer says he built a hand-drawn video engine in one evening, then revised it after the first full video came out slow and visually static. In an account published November 6 on ThorstenMeyerAI.com, he describes measuring the shortcomings, changing the engine and turning the measurements into checks that can stop weak scenes before rendering.
Meyer says the initial concept was inspired by a video style he liked, but he set a clean-room rule: record only general traits, such as hand-drawn strokes and a continuous illustrated world, and use no drawings, assets, layouts or drawing code from the original. He says Claude built the engine, using his own design system and an openly licensed handwriting font. OpenAI’s Codex reviewed the code rather than building it. Meyer reports that Codex found five issues in the first review round; each fix was paired with a test designed to fail against the earlier code.
A roughly one-minute test video received Meyer’s approval and went live that evening in an article about local votes against data centers. The next morning’s first full video did not meet his expectations: he said it had too much empty white space, icons that did not support the narration, and images that stopped moving once drawn. Meyer says that video remains online unchanged as an example of what went wrong.
Instead of relying only on subjective tweaks, Meyer says he measured frame coverage and the time images remained still. Across the first three scenes of a rebuilt video, he reports, filled-frame coverage rose from 15% to 66%, still time fell from 29 seconds to zero, and camera moves increased from three to 27. He says these measures now act as production gates: an overly empty scene can stop before rendering, while an object that does not illustrate the narration prompts a warning.
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 Metrics to Production Gates
The account illustrates a practical distinction for teams using generative tools: a system that can produce a working demo may still make output that is not fit for publication. Meyer’s example places human approval, independent code review and checks on the finished scenes between generation and release, rather than treating a successful run as proof of quality.
The reported numbers also show how a creative judgment can be translated into operational tests. Frame coverage and still time do not establish whether a video is accurate, clear or engaging, but they can flag specific visual problems Meyer identified. His approach makes those measures part of the build process, so the system can catch some recurring flaws before a person reviews the final result.
The evidence remains a single creator’s account, not an independently evaluated comparison or a general benchmark for AI video tools. Readers can take the production practices as an example, but the reported improvements do not establish that the same workflow will produce similar results elsewhere.
Popular right now
Hisense 40-Inch Class QD4SR Series Hi-QLED FHD Smart Roku TV (2026 Model)
Ninja CREAMi 7-in-1 Ice Cream Maker, NC301
Ninja Flip Toaster Oven and Air Fryer, 8-in-1, Stainless Steel, SP151
As an affiliate, we earn on qualifying purchases.
How the First Video Changed
Meyer frames the project as a record of moving an AI idea into production, with failures included. The initial one-minute test and the first full video were separate milestones: he says the test went live the evening he built the engine, while the weak full video appeared the following morning. His subsequent redesign addressed the issues he had noted in that longer piece.
The article also describes a separate experiment from the same week. Meyer says Claude edited the weights of a local model for a legal-answer task. The model completed 12 of 15 answers in the benchmark, compared with four before the edit, but its legal score remained zero. He presents this as another warning that more completed answers did not mean better legal performance. The source does not provide enough benchmark detail to assess the test independently.
„And honestly, it was slow and boring.“
— Thorsten Meyer, describing the first full video
What the Report Does Not Establish
The figures come from Meyer’s own build notes, incident notes and experiments, as described in his article; no independent audit or outside review is supplied. The source gives results for the first three scenes of the rebuilt video, not a broader sample, and does not specify the exact calculation method for frame coverage or the conditions used to count camera moves.
It is also unclear how reliably the new gates identify whether an object genuinely supports narration, or how often they stop a scene that a human would accept. Meyer reports that the improved-look video went live a few hours after his critique, but the account does not give independent audience data or a comparison of viewer response. The legal-model experiment likewise lacks enough detail here to judge the benchmark’s design or the significance of its zero score.
How the New Checks Will Be Used
Meyer says the measurements now operate as gates in the video build: a scene judged too empty is stopped before rendering, and a mismatch between an object and the narration triggers a complaint. He says the video shown with his article was made using the revised approach. The source does not announce a publication date for another test or a formal evaluation of the gates.
The next observable step is whether the checks continue to catch the problems they target in later videos, and whether human review remains part of approval. For now, the published account documents one creator’s process and reported results; it does not establish how the workflow performs across other subjects or production teams.
Key Questions
What did Meyer build?
He says he built an engine for a hand-drawn, continuously illustrated video style in one evening, with Claude writing the code and OpenAI’s Codex reviewing it.
What was wrong with the first full video?
Meyer described it as slow and boring, citing too much white space, icons that did not support the narration and images that remained still after they were drawn.
What changed in the rebuilt video?
For its first three scenes, Meyer reports frame coverage rising from 15% to 66%, still time dropping from 29 seconds to zero, and camera moves increasing from three to 27. He says these measures became checks in the production process.
Were the results independently verified?
The source presents the numbers as Meyer’s own measurements and does not include an independent audit. The account also does not provide a broader sample or audience-response data.
What happened in the legal-model experiment?
Meyer says a locally run model completed 12 of 15 legal benchmark answers after a weight edit, compared with four beforehand, while its reported legal score stayed at zero. The article excerpt does not provide enough benchmark detail for an independent assessment.
Source: ThorstenMeyerAI.com
