People ask how one person keeps a channel moving, so I made a video walking through my whole setup. Here is the gist. The boring parts stopped being manual. They became custom skills inside a seven-stage Claude Code pipeline that runs from my terminal: Strategy, Ideate, Produce, Publish, Measure, Learn, then Loop back to Strategy.
Strategy lives in the system prompt
Stage one is my CLAUDE.md, the file Claude reads every session. It holds what actually works on my channel, from real performance data, not a guess. For example, topic videos without a specific tool get about 11 views, while tool-launch videos with a $0 angle pull 200-plus watch hours. Title formulas, hook structures, and production rules sit in the same file, so when I ask what to record, Claude ranks options against a proven formula instead of brainstorming randomly.
Ideate and produce
For ideate, I run a workflow that pulls the last 30 days of discussion across YouTube, Reddit, X, Hacker News, and the web and synthesizes it. I also screenshot YouTube search results for a topic and paste them in, and Claude reads the packaging, color, text, and framing. Then I check for a first-mover window: a tool that just launched with zero tutorials yet.
Produce is the longest stage and happens before I hit record. Claude generates ten-plus title options and scores each on six dimensions: immediate hook, measurable outcome, personal element, audience clarity, curiosity, and urgency. I pick the top three for testing. Thumbnails get a design brief, not an image, covering focal point, emotion, and text size. Scripts use my hook framework for the first 30 seconds, and slide decks generate from one command. Nothing gets recorded until the packaging is done.
Measuring instead of guessing
This is where most creators stop. After publishing, Claude writes the metadata template, then cross-posts to LinkedIn, Twitter, and the web in each platform’s voice. Then measure keeps it honest. I A/B test titles on plateaued videos through the YouTube API: run variant A, measure for 48 hours, swap to B, measure, apply the winner. I also pull my full corpus, every transcript, comment, and description structured as YAML files Claude can search, so I can ask what commenters keep requesting that I have not covered. The answers come from data, not a hunch.
Learn closes the loop. Every session logs what I did and what worked, with a permanent “what’s working versus what’s not” section at the top of the daily log that Claude reads each time. If a video flopped, the reason is logged, and Claude will not let me repeat it. The system gets smarter because the data accumulates, not because I do.
The open-source pieces
You can read the actual setup. The simplified CLAUDE.md is at github.com/ravsau/ai-tutorials/tree/main/claude-code-for-youtubers, and the creator skills pack is at github.com/ravsau/youtuber-skills. Both are open source, so you can lift the skills into your own channel.
Watch the video above for the full stage-by-stage walkthrough, and subscribe if you want more of these AI engineering builds. I specialize in building Claude Code skill pipelines that turn a manual workflow into a system that improves each cycle. Book a call at cloudyeti.io/meet.