Read the full analysis: AI Project Lessons: A First Video, A Long Evening, And A Few Rules on ThorstenMeyerAI.com
TL;DR
Thorsten Meyer says an AI-assisted whiteboard-video project went from an initial idea to a published short test in one evening, but its first full video was slow and visually weak. After a human review and changes guided by measurements, the project added checks for frame coverage, stillness and whether visuals matched the narration. The account is a personal report, and its metrics come from the author’s own notes and experiments.
Recommended by digitechbytes.com · AI-assisted guides
Shopping for technology news & gadgets? Start with the guides we keep up to date:
Updated November 202615 Best Video Switchers for Live Streaming in 2027See the top picks →
Thorsten Meyer says he built an AI-assisted engine for hand-drawn explainer videos in one evening, but the first full video it produced the next morning was, by his assessment, slow and boring. In an account published on his site and updated November 9, he describes how a second AI reviewed the code, a human approved a short test, and later production checks turned visual criticism into measurable rules.
Meyer describes the project as a personal account of taking AI ideas into production, including failures. He says the initial inspiration was a hand-drawn explainer video, but his team set a “clean room” rule: record a short list of broad style traits, then build its own engine without using the source video’s drawings, assets, layouts or code. The project used Meyer’s design system and an openly licensed handwriting font.
According to Meyer, Claude built the engine and OpenAI’s Codex reviewed the code without contributing to the build. Codex raised five findings in its first review, which Meyer says were fixed that evening. Each finding received a test designed to fail against the earlier code. He also says he watched a roughly one-minute test video and approved it before it was published inside an article about local votes against data centers.
The first full video, produced the following morning, prompted a different response. Meyer says it had too much white space, icons that did not support the narration, and images that stopped moving once drawn. After measuring the first three scenes, he reports that frame coverage rose from 15% to 66%, stillness fell from 29 seconds to zero, and camera moves increased from 3 to 27 in the rebuilt version. These are the author’s reported figures for those scenes, not an independently verified evaluation.
Build notes
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 layoutsSame eveningOwn engine, first reviewClaude builds; OpenAI’s Codex reviews; 5 findings fixed, each with a test that fails on the old codeSame eveningA human says yesOne-minute test video approved and publishedNext morningFirst full video: slow and boringToo much white space, icons that did not fit, a picture that stood stillHours 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 Build Checks
The account’s main point is that a working demonstration does not, by itself, show that an AI-generated product is good enough to publish. Meyer describes three distinct safeguards: independent code review, a human approval step, and tests that block production when a result misses defined criteria.
That approach matters to teams using AI to make media or software because a project can appear successful while failing at its intended purpose. Meyer says the video engine now checks whether scenes fill the frame and whether pictured objects relate to what is being said. Those checks make some subjective criticism actionable, although the reported figures alone do not establish that viewers found the revised video more engaging or informative.
Meyer also connects the video project to a separate legal-model experiment. In that test, he says a model answered all 10 published tasks, up from four, while its legal score remained zero. The result illustrates a narrow but important distinction in his account: completion is not correctness. A model can produce more answers without improving the measure that matters.
How the Video Engine Was Revised
The initial one-minute test and the first full video were separate stages. Meyer says the short test was approved and published the evening the engine was built. The next morning’s full video was left online unchanged as an example of what the team considered unsuccessful. He says the improved version, about the Nasdaq’s worst day in 19 months, went live a few hours after his critique.
For the revision, the team measured two things: how much of the frame contained content and how long the image remained completely still. Meyer says the engine was then adjusted so shots filled more of the frame, camera movement tracked sentences, completed drawings kept moving slightly, and objects illustrated something in the narration. The reported measurements for the first three scenes were then turned into gates: a scene judged too empty could be stopped before rendering, and an irrelevant visual could trigger a build warning.
The post was updated November 9 to say that the figures refer to 10 published Lexassay tasks and that the runs are public on lexassay.com. Meyer says the figures come from his build notes, incident notes or published experiments. The available source is his own account rather than an independent audit of the project.
“A demo that runs is not a decision.”
— Thorsten Meyer
What the Report Does Not Establish
The account does not provide an independent assessment of the videos, the code-review findings or the measurements. It also does not specify the measurement method in enough detail to assess how frame coverage, stillness or camera moves were counted. The reported change covers the first three scenes, not necessarily every scene in the finished video.
It is also unclear whether viewers responded better to the revised video, whether the new checks catch all relevant quality problems, or how often they block a build. Meyer’s report describes the project’s rules and his own results, but it does not establish that the same process will produce similar outcomes for other teams. The legal benchmark result is likewise limited: the account reports more completed tasks but a legal score of zero, without claiming that the model became legally reliable.
The Checks Move Into Production
Meyer says the video measurements now act as production gates, rather than remaining observations from a one-time review. The next practical test is whether those gates consistently catch empty scenes and visuals that do not support the narration before a video is rendered. His account does not give a schedule for further evaluations or say whether additional benchmarks will be published.
The reported Lexassay runs are public, according to the November 9 update. Readers can compare the stated task-completion results there, while the broader claims about video quality and the effectiveness of the review process remain based on Meyer’s account. Further results would help show whether the project’s checks improve later videos, not just the version described in the post.
Key Questions
What happened in the AI video project?
Thorsten Meyer says he built an AI-assisted engine for hand-drawn explainer videos, approved a short test, then rejected the first full video as slow and visually weak. He describes revising the engine and adding production checks based on measurements.
Which AI tools did Meyer say were used?
Meyer says Claude built the engine and OpenAI’s Codex reviewed the code. He says Codex did not build the project and that five findings from its first review were fixed with tests.
What changed in the revised video scenes?
For the first three scenes, Meyer reports that frame coverage rose from 15% to 66%, stillness dropped from 29 seconds to zero, and camera moves increased from three to 27. The account does not independently verify these figures.
Does the legal-model experiment show the model became better at law?
Not based on the figures reported. Meyer says the model answered all 10 published tasks instead of four, but its legal score stayed at zero. The source presents task completion and legal performance as different measures.
Source: ThorstenMeyerAI.com