🔍 Read the full analysis: AI To Production: What I Learned By Making The First Video 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 a hand-drawn video engine in one evening, but its first full video was slow and visually weak. After measuring frame use, stillness and camera movement, he added checks to the workflow; the account also describes independent AI code review and human approval.
Thorsten Meyer says he built a hand-drawn video engine in one evening, but the first full video it produced the next morning was, in his words, slow and boring. In an account published on ThorstenMeyerAI.com and updated November 9, he describes how he used AI code review, human approval and measured checks to revise the workflow before publishing a stronger video, offering a practical account of the gap between generating a demo and putting AI-assisted work into production.
Meyer says the initial idea came from a hand-drawn explainer video he liked. To avoid copying it, he says he wrote down broad style traits rather than reusing its drawings, layouts or other assets. He describes the resulting engine as built from scratch with Claude, using his own design system and an openly licensed handwriting font. The video style used a continuous illustrated world, hand-drawn strokes and large colored numbers.
He says OpenAI’s Codex reviewed the code separately from Claude, which did the building. Codex identified five issues in the first review round, and Meyer says each was fixed with a test that failed on the earlier code. He also describes watching a roughly one-minute test video and approving it before it was published that evening in an article about local votes against data centers.
The following morning’s full video drew a different response from Meyer: he found too much white space, icons that did not support the narration, and drawings that stopped moving once complete. He measured the first three scenes of a rebuilt video. According to his account, frame fill rose from 15% to 66%, still time fell from 29 seconds to zero, and camera moves rose from 3 to 27. He says those measures became checks that can stop a scene from rendering if it is too empty, while another check flags objects that do not illustrate the spoken material.
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 Demo to Quality Gates
The account matters because it shows how Meyer distinguishes a working generation process from a result he considers ready to publish. The first video did render, but his review found problems that a simple completion check would not have caught. By turning specific critiques into measurable checks, he says the workflow can detect some recurring shortcomings before a new video is rendered.
It also presents oversight as part of the production process rather than a final formality. Meyer says one AI system builds while a separate system reviews, and a human decides whether a video is suitable to publish. Those steps do not prove that the output is accurate or good; they describe safeguards in his own workflow. The account’s central lesson is that completion is not the same as quality, and the measures selected still need to match what the work is meant to achieve.
Popular right now
DREAME A3 AWD 1000 LiDAR Robot Lawn Mower
Roborock RockNeo Q110H Wire-Free Robot Lawn Mower
Apple AirPods Pro (1st Generation) with MagSafe Charging Case
As an affiliate, we earn on qualifying purchases.
The First Cut and Its Revision
Meyer frames the article as a personal account based on his build notes, incident notes and experiments, rather than an independent evaluation of AI video tools. He says the first short test was approved and published on the same evening the engine was built. The next morning, the first full video made in the new style was left online unchanged as an example of what he did not want to repeat.
His revised video concerned the Nasdaq’s worst day in 19 months, according to the article. Meyer says it went live a few hours after he critiqued the first full video. The updated article also refers to a separate experiment in which a model’s answers on 10 published legal tasks improved from four correct responses to all 10 after a weight edit, while its legal score remained at zero. That example reinforces his warning that a higher task-completion count alone can give a misleading impression of performance.
“A demo that runs is not a decision.”
— Thorsten Meyer
Limits of the Published Measures
The article reports results from Meyer’s own workflow, but it does not provide an independent assessment of the videos or the quality gates. The measurements cover the first three scenes of the rebuilt video; the account does not establish that the same changes occurred across every scene or across other projects. It also does not specify how frame fill, stillness or whether an object supports narration are calculated in every case.
The source describes the legal-task result but says the model’s legal score remained at zero. It does not provide enough detail here to assess the benchmark, the scoring method or why the task answers and legal score diverged. More broadly, the account does not report a measured comparison of production time, costs or audience response before and after the revised checks.
Checks for Future Video Builds
Meyer says the new measures now act as gates in the video-building process: an overly empty scene can be stopped before rendering, and an object that does not match the narration can prompt a warning. The article presents those checks as part of his current workflow, but does not give a schedule for further releases or describe a formal external evaluation.
Readers can expect the next useful evidence to come from how the checks perform on later videos: whether they catch recurring problems, whether human review remains necessary, and whether the measures align with the finished work. Meyer also notes that monitoring is useful only if it reaches a person. He recounts that a guard previously produced about 1,300 warnings in a log nobody read, while an alert email failed to send. The source does not say whether that alert failure has since been fully resolved.
Key Questions
What did Thorsten Meyer report?
He described building a hand-drawn video engine, finding the first full video unsatisfactory, and revising the workflow using code review, human approval and measured checks.
What changed in the rebuilt video’s first three scenes?
According to Meyer, frame fill rose from 15% to 66%, stillness decreased from 29 seconds to zero, and camera moves increased from three to 27. These figures apply to the first three scenes, not necessarily the full video or other projects.
Which AI tools did Meyer say he used?
He says Claude built the engine and OpenAI’s Codex reviewed its code. In his setup, Codex checks and reports issues rather than building the system.
Was the first full video replaced?
No. Meyer says the first full video remains online unchanged as an example of the result he considered slow and boring. He says a revised video went live a few hours after his critique.
Do the reported measures prove the videos are accurate or effective?
No. The article describes workflow checks for visual qualities such as frame fill and movement, along with a check that objects illustrate the narration. It does not report an independent evaluation or show that those measures prove factual accuracy or audience effectiveness.
Source: ThorstenMeyerAI.com
