🔍 Read the full analysis: I Tried To Take An AI Idea To Production In One Evening on ThorstenMeyerAI.com
Get business pricing on office and shipping supplies
- Business-only prices and quantity discounts
- Tax-exempt purchasing
- Multiple users, one account, clear invoices
TL;DR
Thorsten Meyer says he built and published a new hand-drawn video style within hours, then found that its first full video was slow and visually weak. He describes revisions guided by measurements, independent AI code review, human approval and checks that block some issues before rendering.
Thorsten Meyer says he turned an idea for a hand-drawn, continuously moving video style into a live tool within hours, but the first full video made with it was slow and visually dull. In an account published November 6 on ThorstenMeyerAI.com, he describes how he used AI code review, human approval and measured production checks to revise the system before publishing another video later that day.
Meyer says he began with a style he admired in another video, then wrote down a limited set of visual characteristics rather than reusing its drawings, assets or layouts. The new engine used his own design system, an openly licensed handwriting font and a hand-drawn aesthetic. He says Claude built the engine, while OpenAI’s Codex reviewed the code without contributing to the build. Five findings from Codex’s first review were fixed that evening, and Meyer says each fix was paired with a test that failed against the earlier code.
A test video of about one minute received Meyer’s approval and went live inside an article about local votes against data centers. The next morning, however, the first full video in the style struck him as too empty and static. He says its icons did not support the narration and completed drawings remained still. Rather than replace that version, he left it online as an example of what the process had produced before revision.
Meyer says the team then measured frame coverage and stillness, and adjusted the engine so shots filled more of the frame, camera movement tracked narration and drawn objects continued moving slightly. On the first three scenes, he reports that frame fill rose from 15% to 66%, stillness fell from 29 seconds to zero, and camera moves increased from three to 27. He says those metrics became checks that can stop a scene from rendering when it is too empty, or flag an object that does not illustrate the words being spoken.
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 Prototype to Production Checks
The account matters less as evidence that an AI project can be built quickly than as a case study in what happened after the first working version. Meyer’s central point is that a functioning demo did not establish that the result was suitable for publication. His next-morning review surfaced problems that were not captured by merely getting the software to run.
The workflow he describes combines different forms of review: one AI system writes code, another looks for defects, tests are designed to expose the earlier behavior, and a human makes the publication decision. The production gates add a further step by turning some subjective feedback—such as excessive empty space—into checks that can interrupt a build. That approach may help teams make creative work more repeatable, although the results reported here are from Meyer’s own project, not an independent evaluation.
His examples also show why a metric needs a defined scope. The video figures cover the first three scenes, not necessarily the full video or later projects. They document changes to those scenes, but do not establish that viewers preferred the revision or that the process reduced production time overall.
Popular right now
Dyson V8 Cyclone Cordless Vacuum
KitchenAid Artisan 5-Qt Tilt-Head Stand Mixer – Dried Rose (KSM150PS)
Breville Bambino Espresso Machine, Brushed Stainless Steel
As an affiliate, we earn on qualifying purchases.
The Project’s Two Early Tests
Meyer frames the video build as one example of a broader production principle: a model or tool can appear to improve on one measure while failing on another. He describes a separate experiment inspired by research on using direct edits to one AI model’s weights to repair another. He asked whether a similar edit could improve one of his local models on legal questions.
According to Meyer, the edited model completed 12 of 15 answers in his legal benchmark, compared with four before the edit, but its legal score remained zero. The account does not provide enough detail in the supplied material to assess the benchmark, scoring method or why completion increased without a higher score. The example supports his caution that completing more responses is not the same as answering them correctly.
Meyer also cites a past monitoring failure from August: a guard reportedly produced about 1,300 warnings in a log that nobody read, while an alert email did not go out. He uses the incident to argue that monitoring has little practical value if it does not reach a person who can respond. The source does not specify the warning period or provide an independent audit of the incident.
“A demo that runs is not a decision.”
— Thorsten Meyer
Limits of the Reported Results
The figures come from Meyer’s own build notes, incident notes and experiments, as described in his article; the supplied account does not include an independent review. The video measurements cover only the first three scenes, and it is unclear how frame coverage, stillness and camera moves were defined or measured. No audience data are provided to show whether viewers found the revised video more engaging.
The article also does not specify how often the new gates reject scenes, whether they catch every relevant quality problem, or how the checks perform across other topics and video lengths. Details of the legal benchmark and its zero score are limited in the supplied material. The reported monitoring incident likewise lacks a stated warning period. These examples are evidence of what Meyer says happened in his projects, not proof that the same workflow will produce similar outcomes elsewhere.
How the New Workflow Will Be Used
Meyer says the improved video look is already being used in published work, including a video about the Nasdaq’s worst day in 19 months and the explainer accompanying his account. The stated next step is continued use of the production gates: scenes that fail the specified checks are stopped before rendering, while issues involving narration and illustration are flagged for review.
The account does not announce a schedule for further releases, new benchmarks or an external evaluation. The next useful evidence would be whether the checks continue to catch defects in later videos, how often human reviewers must intervene, and whether viewers respond differently to the revised style. For now, the published results describe an early workflow and its initial adjustments.
Key Questions
What did Thorsten Meyer build?
He says he built an engine for videos using a hand-drawn visual style, continuous scenes and camera movement, then used it to make videos for his site.
How did the first full video perform?
Meyer described it as slow and boring, citing empty space, icons that did not support the narration and drawings that stopped moving when complete.
What changed after the first video?
He says the team measured frame coverage, stillness and camera movement, changed the engine, and turned some quality checks into gates that can stop rendering.
Did the reported measurements cover the whole video?
No. The source reports the changes for the first three scenes. It does not say that the same figures apply to the full video.
Does the account prove the workflow works for other teams?
No. The figures and incidents are reported by Meyer from his own projects, and the supplied account does not provide an independent evaluation or audience study.
Source: ThorstenMeyerAI.com
