🔍 Read the full analysis: What My First Production AI Video Taught Me About Starting Small on ThorstenMeyerAI.com
Get smart everyday buys 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 his first full video using a new hand-drawn AI visual style was slow and visually unengaging, despite an earlier short test receiving his approval. He measured the problems, changed the video engine and made the resulting checks part of production; the account is based on his own build notes and experiments.
Thorsten Meyer says a new AI-generated, hand-drawn video style moved from an evening experiment to a published video within hours, but its first full production was slow and visually boring. In an account published November 6, he describes how human approval, code review and measured production checks helped him revise the work rather than treat a running demo as proof it was ready.
Meyer said the idea came from liking another creator’s hand-drawn explainer. To avoid copying it, he says he wrote down a limited set of general traits: hand-drawn strokes, a continuous visual world rather than slides, and large numbers used as color accents. He says the team used no drawings, assets or layouts from the reference video, built its own engine, used colors from his design system and chose an openly licensed handwriting font.
Meyer said Claude built the engine and OpenAI’s Codex reviewed the code without building it. Codex reported five findings in the first review round, he said, and each fix was paired with a test that failed against the earlier code. Meyer describes that as a way to check whether a test can detect the original problem, rather than merely pass after a change.
A roughly one-minute test video received Meyer’s approval and was published within an article about local votes against data centers, according to his account. The first full video, made the following morning, did not meet his expectations: he recorded too much white space, icons that did not support the narration and finished drawings that remained still. He says that version remains online as an example of what the team wanted to improve.
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 Production Checks
The account’s practical point is that getting a video to render is not the same as judging its quality. Meyer says his initial reaction to a short test was positive, while the longer, full production exposed problems that were not apparent in the demo. That distinction matters to readers using generative AI in publishing or other production work: a successful output may still require human review and criteria tied to the intended result.
After critiquing the full video, Meyer’s team measured frame coverage and time spent with no movement. On the first three scenes of the rebuilt video, he reports that filled frame area rose from 15% to 66%, still time fell from 29 seconds to zero, and camera moves increased from three to 27. These are figures from his own project, not independent evidence that the same approach will improve other video systems.
Meyer says the measurements were then turned into checks that can stop a render when a scene is too empty, or flag an object that does not illustrate the narration. The broader lesson he draws is operational: feedback is more useful when it leads to a test or gate that can catch a repeated problem before publication.
Products Worth Considering
Dell XPS 13 9345 Laptop, 13.4-Inch FHD+ 120Hz, Snapdragon X Elite, 32GB RAM, 1TB SSD
Apple MacBook Neo 13-inch Laptop with A18 Pro Chip, 8GB Memory, 256GB SSD, Liquid Retina Display, and 1080p FaceTime HD Camera – Blush
Canon imagePROGRAF PRO-1100 17-Inch Professional Wireless Inkjet Photo Printer
As an affiliate, we earn on qualifying purchases.
How the Video Workflow Developed
Meyer presents the video as a personal account of moving an AI idea into production, not as a controlled comparison or an independently evaluated study. He says the style engine was built in one evening, the short test was approved later that evening, and the first full video was reviewed the next morning. A revised video went live a few hours after that critique, according to his account.
The source also describes a separate experiment involving a local model and a legal benchmark. Meyer says an AI-assisted weight edit increased the number of completed answers from four of 15 to 12 of 15, while the model’s legal score remained zero. The result illustrates his distinction between task completion and correctness, but the source does not provide enough benchmark detail to assess the experiment independently.
Meyer also cites an earlier monitoring failure from August: a guard wrote about 1,300 warnings to an unread log, and an alert email did not go out. He uses that example to argue that monitoring needs to reach a person to be useful. These examples are his reported build notes and incidents, rather than findings independently verified in the article.
“It was slow and boring.”
— Thorsten Meyer, describing the first full video
Limits of the Reported Results
The production measurements are Meyer’s own figures for the first three scenes; the account does not say whether the same changes were measured across a larger set of videos or independently assessed. It also does not provide a detailed scoring method for frame coverage, still time or whether objects appropriately illustrate narration.
The source does not identify the reference video or provide an independent assessment of whether the new work is sufficiently distinct from it. Meyer says no assets or layouts were reused and describes a clean-room approach, but that account is not independently verified here. The legal benchmark’s scoring rules and the weight-edit method are also not set out in enough detail to evaluate the result beyond the figures he reports.
Checks for Future Video Builds
Meyer says the new measurements now function as production gates: a scene that is too empty can be stopped before rendering, while an object that does not illustrate the spoken material can trigger a warning. Future videos made with the system will show whether those checks consistently catch the issues he identified.
The source does not announce a separate release date, a broader rollout or an external evaluation. For now, the next step described is continued use of the revised workflow: code review, tests for identified bugs, human approval and checks against the video’s own production criteria.
Key Questions
What happened to Meyer’s first full AI-generated video?
Meyer says it was slow and visually boring, with excessive white space, unhelpful icons and drawings that stopped moving once completed. He says it remains online unchanged as an example of what the team wanted to fix.
How did the revised video differ?
For the first three scenes, Meyer reports that frame coverage rose from 15% to 66%, still time dropped from 29 seconds to zero, and camera moves increased from three to 27. Those are his project measurements, not an independent evaluation.
What roles did Claude and Codex have?
According to Meyer, Claude built the engine, while OpenAI’s Codex reviewed its code and reported issues. He says each of five findings from the initial review was paired with a test that failed on the earlier code.
Does the legal-model experiment show the model became better at law?
Not based on the result Meyer reports. The model completed 12 of 15 benchmark answers after the edit, compared with four before, but its legal score remained zero. The source does not provide enough benchmark detail for a fuller independent assessment.
Source: ThorstenMeyerAI.com
