Obsidian Webclipper Trial and Error
Obsidian WebClipper YouTube Clipping First Use
Obsidian WebClipper supports the interpreter option and works with Gemini, Claude, and OpenAI, but it also supports Ollama and OpenAI-compatible local LLMs.
When I have it analyze based on transcripts using Gemma4 to create summaries, key facts, and key actions based on the transcript, I get pretty decent results.
What's amazing is that I didn't have to run any special programs—I just set up templates in the web clipper and had a local LLM running, and it was possible. Some of it even works on mobile.
The Problem Starts and Gets Solved
Many YouTube videos don't have transcripts, more than I expected.
While thinking about how to handle this, I discovered something called Whisper.
Extract a script with timestamps using Whisper -> analyze with Gemma4 -> process the markdown file -> add to Obsidian
I had Claude write a Python program to implement this pipeline, and since running it via CLI was tedious, I registered it as a Raycast script and created a Telegram bot to integrate it. This gave me the ability to remotely clip to the Mac at home when I can't use the web clipper.
As a side effect, YouTube transcripts in Korean have lower accuracy than expected, but the scripts extracted with Whisper have quite good quality, so the analysis quality seemed a bit better.
The Trial and Error Begins, But Then...
Just when everything was solved and I was about to sleep peacefully, I noticed there was a "Ask a question" button at the bottom of YouTube videos.
Like I was enchanted, I clicked the button and a chat window opened where I could ask questions and get answers.
The content seemed much better than what Gemma4 produces.
I wondered if I could extract this in the web clipper, so I opened the browser developer tools and tinkered around until I found a way to extract the chat window content and shouted Eureka.
The problem was that the items I was extracting using Gemma4 were three, and to fit that content into the template format, I needed to ask three separate questions.
It wasn't particularly difficult, but the fact that I had to ask questions three times when it was possible with one click was incredibly frustrating.
I could ask all questions at once and insert the content in one go, but that would shake the existing format I was using, which I also hated.
Of course, clipping is a raw data concept and Claude, the Obsidian vault manager, organizes it well when wiki-fying it anyway.
After modifying templates and trying various things, it occurred to me that if I craft the question prompt better, wouldn't the AI answer in the exact format I want all at once?
I remembered from before in the CLAUDE.md document that when emphasizing to AI, if you add guidelines like to the instructions, it follows well, and recently I also heard that AI follows commands well when given in XML-like formats.
So I combined the three questions into one and attached format each item with ## instead of numbers at the top of the question, and it magically added ## in front of each item, successfully achieving the format.
I can't believe I was wrestling with this until 2 AM ㅋㅋㅋ
Conclusion
To be honest, it's something I could do or not do, but it was satisfying and it was a really strange day.
With vibe coding, Gemini CLI, Claude Code, and similar things, I've only properly used them for less than a month, and before that it seemed I only used them as a more convenient search engine than Google or a conversation partner.
Since I'm a freelancer and don't have projects right now, I'm resting at home, and before I would have been rolling around in bed, but now I'm spending several hours a day playing around with AI.
It's not about thinking "I should try because I need this" but rather "let me do it because I can," and as I tried different things, it became fun.
Before, even when I had ideas or things in my head, my mind would go blank and I couldn't even attempt many things, but with AI, attempting becomes easy and the process of ideas becoming concrete is joyful.
ps. I'm sharing a few clipping settings I use, so if anyone needs them, feel free to use them. Of course, you'll need to modify them, and it's difficult to use without LLM connection.
If you have any questions when using them, please leave a comment.
