🔍 Read the full analysis: What A Single Evening Can Teach You About AI Production on ThorstenMeyerAI.com
Get business pricing on monitors, keyboards and dev gear
- 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 AI video engine in one evening, then found its first full video slow and visually static. After measuring the problems, revising the engine and adding automated gates and human approval, he published an improved version; he also described a separate model-editing experiment where task completion rose but the legal score remained zero.
Thorsten Meyer says he turned a hand-drawn video style into a working AI production engine in one evening, but its first full video was, in his words, slow and boring. In an account published on ThorstenMeyerAI.com and updated November 9, he describes how code review by a second AI, human approval, measured revisions and automated checks helped produce a later video he considered ready to publish. The account offers a practical case study in the gap between making an AI demo run and deciding that its output is fit to go live.
Meyer says the idea came from a hand-drawn explainer video he liked. To avoid copying it, he says his team wrote down a short list of general style traits—a hand-drawn line, a continuous visual world and large numbers used as accents—and used no drawings, layouts, assets or drawing code from the original. He says the engine was built with Claude, used a design system of his own, and used an openly licensed handwriting font.
A separate system played the role of reviewer: Meyer says OpenAI’s Codex read the code and tried to find problems but did not build the engine. It identified five findings in its first review round, he says, and each fix was paired with a test that failed against the earlier code. Later, Meyer watched a roughly one-minute test video and approved it for publication; he says it appeared in an article about local votes against data centers that evening.
The first full video produced the next morning did not meet his expectations. Meyer identified empty-looking frames, icons that did not support the narration and drawings that stopped moving when complete. He says the team measured frame coverage and still time, then revised the engine. Across the first three scenes, filled frame rose from 15% to 66%, still time fell from 29 seconds to zero, and camera moves increased from three to 27. He says those measures became build gates, and an improved video went live a few hours after his critique.
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 Demo to Release Decision
The account matters because it frames AI production as a chain of decisions rather than a one-step generation task. A working tool and an output that looks plausible do not, by themselves, establish that the result is accurate, useful or ready for an audience. Meyer’s example shows how a creator can turn subjective criticism—too much blank space or images that do not support the words—into checks that can block a later build.
It also shows the limits of relying on automated review alone. In Meyer’s process, Codex supplied code findings, while a human made the release decision. That division is relevant to teams using AI to create media or software: automated checks can catch specified problems, but approval still depends on people defining what quality means and examining the result. The measurements he reports are specific to his first three scenes, not evidence that all AI video systems will improve by the same amount.
Meyer’s separate legal-model experiment reinforces the distinction between a convenient metric and the outcome that matters. He says a model answered all 10 published legal tasks, up from four, after a weight-edit experiment, while its legal score remained zero. A higher completion count did not translate into a better score on the benchmark’s legal measure.
Products Worth Considering
Samsung Galaxy Tab A11+ 8GB RAM, 256GB Storage, Gray
Jackery Explorer 1000 v2 Portable Power Station, 1070Wh, 1500W
Samsung Galaxy Tab A9+ 11-Inch Android Tablet, 64GB, Graphite
As an affiliate, we earn on qualifying purchases.
Two Tests of AI Quality
Meyer presents the video build as a personal account, drawing on his own build notes and incident notes. The article was dated November 6 and updated November 9, 2026; the update says the figures refer to 10 published Lexassay tasks and that the runs are public on lexassay.com. The reported production metrics concern the first three scenes of the rebuilt video, rather than a broad study or an independent evaluation.
The article connects that production work to a separate experiment prompted by research on one AI repairing another through direct edits to model weights. Meyer says he asked Claude to edit a local model for legal tasks. The model’s task-completion count rose from four to 10, but its legal score stayed at zero. Those outcomes measure different things: whether the model answered a task and whether its answers earned a positive legal score.
Meyer also describes a previous monitoring failure from August. He says a guard wrote about 1,300 warnings to a log that nobody read, while an alert email did not go out. The account does not provide a comparison baseline for that warning count. His point is that monitoring has limited value if a warning is neither delivered nor reviewed.
“A demo that runs is not a decision.”
— Thorsten Meyer
Limits of the Reported Results
The account does not describe an independent assessment of the videos, provide a formal user study, or establish whether the reported visual measures predict audience response. The improvement figures apply to the first three scenes and are reported by Meyer; the article does not give the full measurement method or results across a larger set of videos. It also does not specify how the engine’s gates determine whether an object illustrates what is being said.
For the legal-model experiment, the published account excerpt gives the task count and score but not enough detail to assess the edit’s broader effects, the benchmark’s scoring method or performance on tasks outside the 10 published examples. Meyer’s warning-log figure likewise lacks a stated time window. The source is a first-person account, not an independently verified production study.
Public Runs and Further Checks
Meyer says the runs for the 10 published Lexassay tasks are available on lexassay.com. That makes the reported legal-task results available for inspection, although the account does not say whether additional evaluations or a follow-up report are planned.
For the video engine, the next step described is continued use of the measurements as release gates: scenes that are too empty can be stopped before rendering, and objects that do not illustrate the narration can prompt a build warning. Meyer says the improved video uses the same production approach. The account does not set out a schedule for future versions or report results from a larger series of releases.
Key Questions
What did Meyer build in one evening?
He says he built an engine for producing hand-drawn-style explainer videos, using Claude to write the code and a separately developed style rather than borrowing assets or layouts from the video that inspired the idea.
What changed after the first full video?
After describing the video as slow and boring, Meyer says he measured frame coverage and still time, then revised the engine. In the first three scenes, frame coverage rose from 15% to 66%, still time dropped from 29 seconds to zero, and camera moves rose from three to 27.
Did the legal model improve?
It answered more of the published tasks—10 rather than four—but Meyer says its legal score remained zero. The task count alone therefore did not show an improvement in the benchmark’s legal score.
What remains unverified?
The figures are reported in Meyer’s personal account. The source does not provide an independent review, audience testing, or enough methodological detail to establish whether the video metrics generalize to other scenes or production systems.
Source: ThorstenMeyerAI.com
