← Lab/Entry 05Still runningDraft copy

Content OS

Learn why a creator's content works, then write something original with what you learned.

Notebook page

The original question: Can software work out why a piece of content performs, from its structure alone, and use that to help me write my own?
Hypothesis: What makes content work is mostly structural (the hook, the sequence, the pacing, the pillars), and structure can be measured and borrowed without copying anyone's words.
What was built or tested: A system that imports a creator's public videos, transcribes and analyzes them, builds an evidence-backed profile of how that creator works, and blends chosen traits into a personal "Content DNA" that drives original scripts.
Technical approach: Let the model read and classify, but never let it do the maths. Every number in a profile is calculated in plain code and traceable to the videos it came from.
Outcomes and discoveries: The whole loop is built, from channel URL to a script you can record. It has not yet been proven on real transcripts at scale, because getting them reliably turned out to be the hard part.

Does: Creator analysis, format discovery, transcripts, content pillars, idea and script generation
Starts from: A YouTube channel URL
Built for: One founder who wants an audience without content becoming a full-time job
Status: Built end to end. Waiting on reliable real-world transcripts.

The idea

Most advice about content is generic. I wanted evidence. If a creator I admire keeps getting results, there should be something measurable about how their videos are built, and I should be able to learn from it without imitating them.

The rule for the whole project: learn structure, never copy language. It is not a plagiarism tool, and nothing in it is designed to make you sound like someone else.

What exists

Creator import. Paste a channel URL and it brings in the public videos and their statistics.
Transcripts and analysis. Each video is transcribed and broken down: content pillar, hook style, script sequence, storytelling devices, calls to action, pacing.
Creator profiles. A statistical picture of how a creator works, where every figure links back to the source videos.
Content DNA. Pick the structural traits you want from several creators and blend them into your own.
Research and scripts. Topic discovery, research briefs with sources and counterpoints, and original scripts checked for speakability and for accidental copying.
A feedback loop. Your own published videos feed back in, so the system learns from your results as well.

The intended day looks like this: open the app, see ranked ideas, pick one, edit the script, record.

What I have found so far

Models should not do arithmetic. Letting a model summarize performance invites confident, wrong numbers. Classification goes to the model. Counting goes to code.
The boring step was the blocker. Analysis, profiles and script generation all work. Reliably getting transcripts for real videos is what held everything up, and everything downstream waits on it.
Synthetic tests only prove so much. The analysis passes its own evaluation, but that evaluation was built on synthetic transcripts. Until it runs on plenty of real ones, I do not know how good it actually is.
Originality needs a check, not a promise. Scripts go through an explicit test for copied source material, because "the prompt says don't copy" is not a safeguard.

Where it stands

Active. The full loop is built and I am working on the transcript problem. Whether it produces scripts worth recording is the experiment that has not finished yet.

Sourcecut →