🔍 Read the full analysis: What It Really Took To Get An AI Idea Into 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’s November 2026 account describes turning a hand-drawn video concept into a published drawing engine, with code review, human approval and measurable checks along the way. His examples show why a working demo or green tests alone do not establish that an AI product is useful or correct.
Thorsten Meyer has published an account of how he moved an AI-generated video idea from a same-evening prototype to a tool used in published videos, describing a process that included separate AI code review, human approval and measurable quality gates. The report matters because it documents not just a fast build, but the checks Meyer says were needed to catch weak output and prevent recurring errors.
Meyer says the idea came after watching a hand-drawn explainer video. He wrote down a limited set of style traits—a hand-drawn line, a continuous visual world and large numbers used as color accents—then had Claude build a drawing engine from scratch. He says the team did not reuse the original video’s drawings, assets, layouts or drawing code. The design used his own color system and an openly licensed handwriting font.
In his setup, OpenAI’s Codex reviews code but does not build it. Meyer says Codex reported five findings in the first review, and each was addressed with a test designed to fail against the old code. He argues that this order helps show a test can detect the original issue, rather than merely pass after a fix.
A roughly one-minute test video went live after Meyer watched it and approved it. The first full video produced the next morning did not meet his standards: he described it as slow and boring, with excessive empty space, unhelpful icons and static images. He measured frame coverage, still-image duration and camera movement, then rebuilt the engine. Across the first three scenes, 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. The figures apply to those three scenes, not necessarily the full video.
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
Why Production Checks Matter
The account illustrates a gap between software that runs and a product that performs its intended job. Meyer’s first full video rendered successfully but failed his review; measurements helped turn a subjective reaction into specific changes, and the resulting thresholds became checks that could stop later renders from proceeding when a scene was too empty.
His examples also underline that different checks answer different questions. Code review can flag implementation problems, while a human can judge whether a video is worth publishing. A benchmark’s completion count, meanwhile, may not establish that answers are correct. For readers building AI products, the report offers a practical case for treating evaluation, approval and monitoring as part of production rather than as optional steps after a demo.
Popular right now
Samsung 43-Inch M70H Series 4K Mini LED Smart TV (2026)
Hisense E6SR Series 55-Inch 4K UHD Hi-QLED Roku Smart TV
LG 77-Inch OLED evo C6 4K Smart TV
As an affiliate, we earn on qualifying purchases.
From Prototype to Published Video
Meyer’s account, published on ThorstenMeyerAI.com on November 6 and updated November 9, 2026, is a first-person report based on his build notes, incident notes and experiments. He says the initial video style was developed in one evening, with an approved short version published that night inside an article about local votes against data centers. The next morning’s longer video became the example that prompted revisions.
The same week, Meyer describes a separate experiment involving a local model and legal questions. He says an AI-assisted weight edit raised the model’s answers from four to all 10 published tasks, while its legal score remained zero. He presents this as a warning that task completion alone can be misleading. The account does not provide enough detail here to independently assess the benchmark or the model’s answers.
“A demo that runs is not a decision.”
— Thorsten Meyer
Limits of the Reported Results
The account is Meyer’s own description, not an independent audit of the video engine, its quality gates or the reported experiments. It does not identify the exact thresholds now used to block scenes, explain how frame coverage and motion were calculated, or give results across a larger set of videos. The reported visual measurements are specifically for the first three scenes.
The legal benchmark result also needs careful interpretation: the report says the model answered all 10 published tasks after editing but had a legal score of zero, without providing the scoring rubric or the answers in this account. It is not clear from the material whether those results have been independently replicated, how the tool performs on other tasks, or whether the monitoring and review process has been tested at a larger scale.
How the Gates Will Be Used
Meyer says the measurements have been turned into production gates: checks can stop a render if a scene is too empty, and the build can flag an object that does not illustrate what is being said. He also says the improved visual approach was used for a video about the Nasdaq’s worst day in 19 months. The account does not set out a schedule for further releases or a broader evaluation.
The next useful evidence would be results across more published videos, details about how the gates are calibrated and examples of issues they catch. Meyer also describes an August incident in which a guard generated about 1,300 warnings in a log that nobody read, while an alert email did not go out. He says this showed that monitoring only helps when it reaches a person; whether that alert path has since been changed is not specified in the provided account.
Key Questions
What was the AI tool built to do?
It was a drawing engine for hand-drawn explainer videos. Meyer says it used a continuous visual world, hand-drawn strokes and large numbers as color accents.
What checks did Meyer use before publishing?
He describes AI code review by OpenAI’s Codex, tests written to fail on the old code, and his own review and approval of a test video. Later, measurements of frame coverage and still time became production gates.
What changed in the rebuilt video?
For the first three scenes, Meyer reports that filled frame area rose from 15% to 66%, still time fell from 29 seconds to zero, and camera moves rose from three to 27.
Did the separate legal-model experiment show that the model became better at law?
Not based on the reported figures alone. Meyer says the model answered all 10 published tasks, up from four, but its legal score remained zero. The scoring details are not included in this account.
Are the results independently verified?
The material presents them as Meyer’s own build and experiment results. It does not describe an independent audit or replication, and several evaluation details remain unspecified.
Source: ThorstenMeyerAI.com
