I run a small YouTube character channel, Cynic Stick, on n8n. This is a walkthrough of how it is wired, with the parts I think are reusable. Much of it was built with Claude through the n8n MCP server: I direct, it builds.
The daily and weekly pipeline (about 25 active workflows)
- Research: scheduled trend and signal searches feed a research workflow. An Ollama research workflow is the fallback.
- Production: one workflow writes and renders the daily short. The weekly long video is split into script and scene prompts, scene render, beat render, compile and review.
- Model registry: one small workflow returns a fixed set of models (writer, medium, lite, image, speech-to-text, text-to-speech). Every other workflow asks it, so a model change is one edit.
- Publishing: separate workflows for YouTube, X, Instagram and Facebook on a schedule. TikTok uploads to drafts and sends me the caption on Telegram for a one-tap manual post.
- Watchdog: 30 minutes before the publish time, a workflow checks that the build started and finished. It messages me only when it did not, so silence means healthy.
- Housekeeping: a daily workflow sorts published shorts into 5 YouTube playlists with a light model. A site export and a chat agent keep cynicstick.com and its “Ask Cynic Stick” widget in sync with the channel.
The new part: a live stream where chat steers the story (12 active workflows)
Viewers type !scene plus an idea (or !ask plus a question) in the YouTube live chat. The stream is an ASCII-art world drawn by a web page that reads an n8n endpoint.
- Poller: reads the live chat on a configurable interval and stays idle unless the automation is on and not locked.
- Ingest: de-duplicates, allows one idea per slot, then calls the moderation sub-workflow.
- Moderation: name and rules checks, then two text judges and three art judges through OpenRouter with strict JSON output. Anything unclear is rejected (fail closed). A model turns the idea into a validated scene, and its output is checked again.
- Human veto: every accepted idea goes to a Telegram message with a veto button. When the window ends, a release workflow closes the message and announces the idea in the chat. Repeated vetoes lock the automation.
- Safety nets: a kill switch (panic and resume webhook), an hourly canary that sends four fixed lines through the judges and locks the automation if a clearly bad one passes, a keeper that checks every minute that the broadcast is live, and a daily retention job that removes raw comment text after 7 days and names and channel IDs after 30.
What I would like feedback on
- Is an hourly canary enough to trust fail-closed judges, or do you test moderation differently?
- How do you balance chat polling latency against YouTube API quota?
- Anything in the retention approach you would change?
Channel: https://www.youtube.com/@cynicstick
Site: https://cynicstick.com
The live is brand new and the audience is tiny, so I care about the design more than the numbers.