🔍 Read the full analysis: How To Move An AI Idea From Concept To Production on ThorstenMeyerAI.com
Get the little things that make your day 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 describes how he turned an idea for a hand-drawn AI video style into a live production workflow, then improved it after the first full video fell short. His account highlights code review by a second AI, human sign-off and measurable checks; it is a personal case study, not evidence that the same results will generalize to other projects.
Thorsten Meyer says he took an idea for a hand-drawn AI video style from concept to a live test in one evening, then revised the production system after its first full video seemed slow and empty. In an account published November 6, he describes using separate AI code review, human approval and measurable quality gates to catch problems before later videos were made.
Meyer says the initial idea came from a video whose style he admired. To avoid copying its specific creative work, he says he wrote down broad traits—such as hand-drawn lines, a continuous visual world and large colored numbers—and built a separate engine without using the source video’s drawings, assets, layouts or drawing code. He says the engine used his own design system and an openly licensed handwriting font.
In his workflow, Claude built the engine and OpenAI’s Codex reviewed the code. Meyer says Codex reported five findings in its first review, which were fixed that evening. He says each finding received a test designed to fail against the old code before the fix was applied. A human review followed: Meyer watched a roughly one-minute test video and approved it for publication, he writes.
The next morning, the first full video in the new style did not meet his expectations. Meyer describes it as slow and boring, citing empty space, icons that did not support the narration and drawings that stopped moving when complete. He measured the first three scenes of a revised video: frame fill rose from 15% to 66%, still time fell from 29 seconds to zero, and camera moves increased from 3 to 27. He says these measures became checks that can stop a scene from rendering if it is too empty.
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
Turning Video Critique Into Checks
The account offers a concrete example of the work between a functioning AI demo and a tool that can be used repeatedly. Meyer’s central point is that a system completing a task does not establish that its output is useful: the first full video ran, but he found it unsatisfactory. Human judgment identified the problem; measurements then made parts of that judgment testable.
His process also separates roles that are sometimes combined in AI development. One model produces code, another reviews it, tests check specific findings, and a person decides whether the result should go live. Meyer’s example does not prove that this arrangement prevents errors in other projects, but it shows how a creator can make review and approval explicit rather than treating a successful run as sufficient evidence.
The production gates matter because they carry lessons from one review into later builds. A check for excessive empty space can flag a problem before a new video is rendered, while a human still has to judge whether the visuals communicate the narration. The approach is relevant to teams using generative tools because it connects subjective quality concerns to repeatable checks without claiming that metrics alone can determine whether a creative product works.
Products Worth Considering
Therabody Theragun Mini 3rd Generation Massage Gun, Black
Fitbit Charge 6 Fitness Tracker – Obsidian/Black
Bowflex SelectTech 840 Adjustable Kettlebell
As an affiliate, we earn on qualifying purchases.
From Style Reference to Original Engine
The source is Meyer’s first-person account on ThorstenMeyerAI.com, dated November 6, 2026. It describes a particular build and the author’s own notes, experiments and incident records, rather than a controlled study or independently audited evaluation. The figures in the article refer to the first three scenes of the revised video, not to a broad sample of videos.
Meyer also connects the video work to a separate experiment with a local model and legal questions. He says Claude edited the model’s weights and that it completed 12 of 15 answers in a legal benchmark, compared with 4 before. However, he says its legal score remained at zero. The example underscores the difference between completing more answers and achieving a better result on the measure that matters; the account does not establish that the edit improved legal capability.
In another example, Meyer says a review round had 12 of his tests passing while four of six checks from the reviewer failed. He also recounts an August incident in which a guard wrote about 1,300 warnings to a log no one read and an alert email did not go out. These are his reported experiences, not independently verified findings, but they explain why he treats test coverage and human-visible monitoring as separate parts of production readiness.
“A demo that runs is not a decision.”
— Thorsten Meyer
Limits of One Creator’s Results
The article provides no independent assessment of the video quality, code review or benchmark results. It does not give enough detail about the legal benchmark’s scoring method, the model, or the weight-edit procedure to determine why the score stayed at zero or whether the result would hold in other tests.
The video metrics cover three scenes in one revised project. Meyer does not report results across a larger video sample, audience response, production costs or whether the gates have prevented later defects. The account also does not establish that a second AI reviewer will reliably identify bugs; its own example of failed reviewer checks suggests that review findings still need to be tested and evaluated.
Applying the Gates to Future Videos
Meyer says the revised checks now stop scenes that are too empty before rendering and flag objects that do not illustrate the narration. The next practical test is whether those gates continue to catch unwanted output across future videos and whether human reviewers find the finished work effective.
His account does not announce a broader product launch, formal evaluation or independent audit. For now, the reported development is a creator’s production workflow and a set of safeguards used on his own site. Further evidence would be needed to establish how consistently the method performs beyond these examples.
Key Questions
What happened in Meyer’s AI video project?
Meyer says he built an engine for a hand-drawn video style, published a short approved test, then revised the system after the first full video fell short.
How did he review the code?
According to Meyer, Claude built the engine and OpenAI’s Codex reviewed it. He says five findings were fixed and each received a test that failed against the old code.
What changed in the revised video?
For the first three scenes, Meyer reports that frame fill increased from 15% to 66%, still time dropped from 29 seconds to zero, and camera moves rose from 3 to 27.
Do the results prove this workflow works for other AI projects?
No. The figures come from Meyer’s own project and notes, and the article does not provide an independent evaluation or results across other teams and systems.
What did the legal-model experiment show?
Meyer says the edited model completed 12 of 15 benchmark answers, up from 4, while its legal score remained zero. The account does not provide enough benchmark detail to interpret the result beyond that report.
Source: ThorstenMeyerAI.com
