Part 2 of a series on wiring my Obsidian vault into Claude. Part 1
In the first post I said the system splits into two halves: the capture-and-batch side that pulls stuff in and tidies it, and the interactive side where Claude and I work in the vault together. This post is about the first half — specifically, getting content in.
Here's the belief the whole thing rests on: the moment you capture something and the moment you organize it should never be the same moment. Every time I've made capture depend on also filing the thing correctly, capture loses. The thought evaporates while I'm deciding which folder it belongs in. So the job I handed to Claude is exactly that deciding — where a thing goes, what it's called, how it's tagged — which frees capture to be as dumb and frictionless as possible.
That means the vault needs more than one front door. Here are the three I use most.
Front door 1: Just tell Claude
This is the one from the origin story in Part 1, except the copy-paste step is gone now. When a half-formed thought shows up — a project idea, a decision I want to remember the reasoning behind, the useful two sentences that fell out of a longer conversation — I just tell Claude to put it in the vault. It picks the folder, writes a real note with proper frontmatter, and commits it.
This has quietly become my single biggest source of new notes. It's the path I reach for whenever I'm already in a conversation and something worth keeping surfaces. (How Claude actually reaches into the vault to do this is the subject of the next post — here I just care that it's the fastest way in.)
Front door 2: Voice, for when my hands are busy
Some of my best thinking happens when I absolutely cannot type — driving, or out walking the dogs on the acreage. Typing-based capture just doesn't exist in those moments, so those thoughts used to die on the vine.
The voice pipeline fixes that. I record a quick memo on my phone, and it's automatically uploaded to the homelab — no button to push after the fact, no remembering to sync it later. From there, a self-hosted faster-whisper instance transcribes it (no audio ever leaving my network); Claude reads the transcript, figures out what it actually is, and structures it into a note; and it commits like everything else. Raw, rambling audio in one end — a filed, titled, tagged note out the other.
This one took the longest to get right, but it's finally working reliably, and it's the piece I didn't know I'd use as much as I do.
Front door 3: Read-later, for other people's content
The first two doors are for my thoughts. This one's for everyone else's. I've long thrown interesting links into Wallabag, my self-hosted read-later app — which, left alone, is really just a nicer-looking graveyard of tabs I meant to read.
So a Node-RED pipeline watches for new saves, hands the article to Claude for a summary, and drops the result into the vault as a resource note. The saved article stops being a someday-maybe and becomes something searchable, summarized, and linkable alongside my own notes. The read-later list finally pays rent.
The obligatory tech debt
I'd love to tell you all three of these run on one clean, unified pipeline. They do not, and the reason is very on-brand for this blog.
I got ambitious and decided I'd consolidate everything onto n8n — one orchestration layer to rule them all. As part of that, I tried to be clever and merge two flows into a single workflow: articles going into the vault and recipes going into Tandoor, my recipe manager. They seemed similar enough — save a link, have something smart parse it, route the result somewhere. Turns out "route the result somewhere" was carrying a lot of weight in that sentence, and the combined workflow ran into enough friction that I set it down.
So for now, the read-later pipeline is still happily running on Node-RED, the n8n migration is a "someday," and I have a half-built workflow sitting in a corner judging me. Creating tech debt for future generations, as promised.
Same shape, every door
Type it, say it, or save someone else's link — underneath, all three doors are the same shape from Part 1: external data → Claude → GitHub commit → human review. The front door changes; the pattern doesn't. And because Claude handles the structuring, I get to stay in capture mode without ever context-switching into librarian mode.
Next up: the other half of the brain — how Claude actually reaches into the vault to read and write it live, mid-conversation. That's where the copy-paste step went to die.
Until then — happy capturing.
This post was written with the help of Claude — fittingly, using the very system it describes.