🔍 Read the full analysis: How I Took An AI Idea To Production In One Evening on ThorstenMeyerAI.com
Get business pricing on tech for your team
- 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 video engine in one evening and published a short test video after reviewing it himself. The next morning’s full video fell short; measurements, code review and automated checks helped improve later videos. These are the author’s account and project-specific results, not an independent evaluation.
Thorsten Meyer says he took a new AI-assisted, hand-drawn video style from idea to a short published test in one evening, then revised the system after its first full video disappointed him the next morning. In an account published on ThorstenMeyerAI.com and updated November 9, Meyer describes a workflow combining AI-generated code, review by a second AI, human approval and measurement-based checks. The account matters as a practical case study in the work between a functioning demo and a production result, though the figures come from Meyer’s own notes and experiments.
Meyer writes that the idea began after he watched a hand-drawn explainer video whose style he liked. To avoid copying it, he says he recorded only broad traits: hand-drawn strokes, a continuous visual world, and large numbers used as colour accents. He says the engine was built from scratch with Claude, using his own design system and an openly licensed handwriting font, rather than borrowing the original video’s drawings, assets or layouts.
In Meyer’s workflow, OpenAI’s Codex reviewed code written with Claude but did not build it. Codex identified five issues in the first review round, he says, and each fix was paired with a test that failed on the old code. Meyer then watched a roughly one-minute test video and approved publishing it within an article about local votes against data centers. He emphasizes that a running demo was not, by itself, the reason to publish: a responsible person watched and approved the result.
The first full video in the new style was less successful. Meyer describes it as slow and boring, citing empty space, icons that did not support the narration and drawings that stopped moving once completed. He measured frame coverage and stillness, then changed the engine. For the first three scenes of the rebuilt video, his figures show filled-frame coverage rising from 15% to 66%, stillness falling from 29 seconds to zero, and camera moves increasing from 3 to 27. Meyer says those measurements became automated gates that can stop scenes that are too empty or flag objects that do not illustrate the spoken content. These are results for the scenes he measured, not an independent benchmark of video quality.
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 Reviewed Video
The account offers a concrete example of why a rapid AI-assisted build is not the same as a reliable production process. Meyer’s timeline was short, but his own description includes several distinct decisions: defining what could be borrowed as inspiration, checking generated code, watching the output, identifying specific defects and turning those defects into repeatable checks. That distinction is useful to readers evaluating claims that an AI tool can take an idea directly to a finished product.
The workflow also highlights limits in automated review. A second model can find software issues, and a metric can flag an overly empty frame, but Meyer’s account still relies on a human to judge whether the test video should go live. The reported before-and-after figures show that selected visual properties changed; they do not establish that viewers preferred the result or that the system performs well across other projects. The practical takeaway is narrower: define what quality means, test the failure state, and retain human approval where the decision depends on judgment.
Products Worth Considering
Samsung Galaxy Tab S7 11-inch Android Tablet 128GB Wi‑Fi with S Pen Mystic Black
LG UltraGear 25G590B-B 25-Inch Full HD 1000Hz Gaming Monitor
Apple AirPods Pro 3 Wireless Earbuds with Active Noise Cancellation
As an affiliate, we earn on qualifying purchases.
A Second Test of AI Evaluation
Meyer connects the video work to a separate experiment he describes from the same week. In that test, he says Claude edited a local model’s weights to improve performance on legal tasks. The model answered all 10 published tasks, compared with four before the edit, but its legal score remained zero. The article treats that result as a warning against equating task completion with correctness. Meyer does not provide enough detail in the supplied account to independently assess the benchmark, scoring method or weight-edit procedure.
He also cites an earlier monitoring failure from August: a guard generated about 1,300 warnings in a log that nobody read, while an alert email did not go out. The figure is a raw count; the account gives no comparison baseline or time window for interpreting it as a rate or increase. Together, these examples support the article’s central point that checks matter only when they test meaningful outcomes and reach someone able to act.
““And honestly, it was slow and boring.””
— Thorsten Meyer, describing the first full video
Limits of the Reported Results
The video measurements are Meyer’s own results for the first three scenes of a rebuilt video; the account does not identify an independent reviewer or report viewer testing. It is also unclear how the frame-coverage and stillness measures were defined in detail, how often the new gates produce false alarms, or whether they have been tested across a larger range of videos. The account names the AI tools used but does not give model versions or enough technical detail to reproduce the build.
The separate legal-model result is also limited by the information provided: the model answered all 10 published tasks, but its legal score stayed at zero. The score’s rubric and the reason for the mismatch are not explained in the supplied material. Meyer’s examples therefore document his workflow and reported outcomes; they do not establish general performance claims about AI coding, video generation or legal reasoning.
Whether the Gates Hold Up
Meyer says the visual measurements now function as build gates: a scene that is too empty can be stopped before rendering, while an object that does not match the narration can trigger a warning. The updated article says the runs for the 10 published Lexassay tasks are public on lexassay.com. The account does not set out a future release date or a formal evaluation plan.
The next useful evidence would be whether the gates catch recurring problems across more videos, whether human reviewers continue to approve the results, and whether the legal benchmark’s score can improve alongside its completion rate. Until then, the published figures are best read as a record of one creator’s development process, including its failures, rather than proof that the approach will generalize.
Key Questions
What did Meyer build in one evening?
He says he built an engine for producing a hand-drawn explainer style with Claude, then published a roughly one-minute test video after watching and approving it.
What changed after the first full video?
Meyer measured frame coverage and stillness after judging the first full video too slow and visually empty. He says the rebuilt video’s first three scenes showed coverage rising from 15% to 66%, stillness dropping from 29 seconds to zero, and camera moves increasing from three to 27.
Did Codex create the video engine?
No. Meyer says Claude built the engine and OpenAI’s Codex reviewed the code, reported issues and helped identify fixes. He describes the tools’ roles as separate.
Do the results prove the video system works generally?
No. The figures come from Meyer’s own build notes and cover the first three scenes of one rebuilt video. The account does not report independent review or broader viewer testing.
What happened in the legal-model experiment?
Meyer says a model answered all 10 published legal tasks after a weight edit, compared with four previously, while its legal score remained zero. The supplied account does not explain the scoring rubric or resolve that discrepancy.
Source: ThorstenMeyerAI.com
