Hey everyone,
I use YouTube for a lot of my research. It is one of the best places to find someone walking through an architecture or a technique in depth. The problem is I do not have time to sit and watch, and a 40 minute talk is rarely 40 minutes of signal.
So when I find a topic I want to go deeper on, or an architecture I want to mimic and start playing with, I do not want to watch the video. I want the information out of it and in front of me in a form I can actually use. Something digestible for me, and just as importantly, something my agents can read too.
I built this into my own app first, but I know not everyone is going to run that. So I pulled it out into a standalone n8n workflow anyone can import. It needs one small Docker service alongside it, and I will explain why below.
What it does
You paste a YouTube link into a form. The workflow pulls the video's captions, runs the transcript through a model, and saves a fully structured markdown research document to Google Drive. Not a two-line summary. A deep-dive breakdown: thesis, step-by-step walkthrough, every command and config value quoted verbatim, a table of the numbers and benchmarks, tools and people named, direct quotes, gotchas, and how to reproduce it on your own stack. The kind of note you can hand straight to another AI system.
How it is built
- Form Trigger takes a single field, the YouTube URL.
- HTTP Request node ("Fetch Transcript") calls the small captions service and gets the transcript back.
- Code node ("Build Prompt") assembles the system and user prompt, dropping the transcript in and asking for the deep-dive structure. This is where the shape of the output lives, so it is easy to tune.
- AI Agent node runs the prompt. The chat model plugs into it as a sub-node.
- Code node ("Build Markdown") wraps the model's output in YAML frontmatter (title, source, channel, video id, date) and builds the filename.
- Google Drive node writes the
.md, and a completion screen hands back the link.
Every node has a sticky note on the canvas, and there is a full sticky walkthrough for the Docker service too, so the workflow documents itself when you import it.
The one Docker service, and why
This is the part that is not pure n8n, so I want to be upfront about it. To transcribe any public video you need the captions YouTube already serves, and the official Data API only lets you download captions for videos you own. The tool that gets around that is yt-dlp, and running yt-dlp inside n8n is awkward, and not possible at all on n8n Cloud. So the template ships a tiny HTTP service that wraps yt-dlp: one endpoint, GET /transcript?url=..., returns the text. It is a small FastAPI app, captions only, no Whisper and no GPU. You build it from the included Dockerfile and run it on the same Docker network as n8n so the workflow can reach it by name. The repo has the service and a walkthrough.
The model is a swappable node
I used n8n's AI Agent node on purpose, because most people are already familiar with it, and because it makes the model a plug. The template ships three chat models wired up as alternates (Claude Sonnet, GPT, and a local Ollama model), so you can run the same transcript through each and compare the output, the quality, the speed, and the cost. An agent uses one at a time; you just drag a different model onto it. Point it at a cheap fast model for a quick summary, or a stronger one for the full deep dive.
Why the output shape matters
The reason I care about the markdown being structured, with frontmatter and consistent sections, is that it has two readers. One is me, skimming it instead of watching the video. The other is my agents. A clean, tagged document drops straight into a RAG store or gets handed to a second agent as context. So the same run that saves me 40 minutes also feeds the systems I have reading on my behalf. That was the whole point: get the knowledge out of the video and into something both a human and a machine can use.
The whole thing is open source (MIT), the workflow JSON, the captions service, and setup docs. github.com/Mfrostbutter/n8n-workflow-templates
If you have built something similar, or you want me to break down any specific node (the prompt builder, the agent wiring, the captions service), I am happy to go deeper in the thread.
Cheers,
Michael