I’m not a writer by birth.
For sixteen years that was the problem I couldn’t solve. I had ideas. I had opinions. I had patterns from working on real systems — decisions made under pressure, incidents that taught me things, architectural instincts built from years of doing it wrong first. But I had no way to express any of it. The knowledge was there. The words never came out right.
I tried to solve this with tools. Lots of them.
The Tool Graveyard
EverNote first. The standard recommendation at the time. Clipping, syncing, notebooks. I used it for a year and accumulated a pile of notes I never looked at again.
OneNote next. Better integration with the Microsoft stack, which was where my work lived. Same result — notes captured, never surfaced, never connected.
Notion after that. The appeal was obvious: flexible, visual, shareable. I built a system that looked great in screenshots. Projects with nested pages, meeting notes linked to action items, a roadmap that updated nicely. But every time I wanted to capture something, Notion’s overhead got in the way — loading the app, navigating to the right database, choosing the right template. For structured work it was fine. For a raw midnight thought, it was too slow. And my notes lived in Notion’s database. Exporting was an afterthought. That felt wrong for something meant to be a lifelong system.
Twos was the closest I’d found. Simple, unstructured, opinionated in the right ways — it shapes itself around how you think rather than asking you to fit a schema. I used it for more than a year. It actually worked for capture.
But after a year I felt the same weight. The same friction, just slower to appear. Notes accumulated. Connections didn’t. The writing problem — the core problem — was still there.
During this whole period I was also experimenting with Obsidian in parallel — trying different folder structures, different tagging systems, different templates. For almost two years. The tool was right. The structure never was. I’d build something, fill it for a few weeks, and then feel the same thing: a pile of notes with no centre of gravity.
At some point I found claudesidian — a PARA + Claude + Obsidian setup someone had built. I cloned it and adapted it to my own work. Getting closer. The DNA was right. But I was still manually driving everything.
Every tool I tried solved the storage problem. None of them solved the expression problem.
The Attempts That Didn’t Work
Before landing on the right setup, I tried two more approaches that seemed promising but failed for specific reasons.
Claude + Slack + GitHub. I gave a Claude app access to my Obsidian KB repository on GitHub — taking notes, updating files, reading content. It worked in principle but had two hard limits. First, it was sandboxed: it couldn’t reach articles I wanted to read and save. I had to run Claude Code locally to fetch and extract anything from the web, which meant my laptop had to be present and open. Second, the Slack integration never quite clicked. The overhead didn’t justify the gain.
OpenClaw (now Clawdbot). I was watching this one for a while. My skepticism about it was confirmed eventually by multiple vulnerability reports and incidents. Ruled it out.
The Night It Clicked
For almost two years I had Obsidian open on my machine and couldn’t make it work. The problem wasn’t the tool. It was that I’d never found a structure that matched how work actually moves.
On January 6th, 2026, I set up Obsidian with PARA. That was the piece that had been missing. Not a new tool — a new structure for the tool I already had. Plain files, the right four categories, notes that lived where they were useful rather than where they were filed. It felt different immediately.
The full system came together on March 9th — the day I set up NanoClaw. I’d found it a few weeks earlier while reading posts on X and sending interesting ones to a WhatsApp group I’d been using to collect resources. One post about NanoClaw stopped me mid-scroll.
I’d been seeing people talk about integrating Claude into Obsidian — plugins, local agents, various setups. But that wasn’t quite the idea I had. My idea was different: a knowledge base accessible from anywhere, with no laptop dependency, no app to open, WhatsApp as the input layer, and an AI that could read and write the files from outside. NanoClaw was built for exactly that architecture. I didn’t find a solution — I recognised that someone had built the missing piece I needed.
But the expression problem — the 16-year one — wasn’t fully solved yet. I still had to open Obsidian. Still had to choose a folder. Still had to write.
NanoClaw launched in early February. On March 9th, I connected it. Lyra — my AI agent — got full access to the KB. That was the night the whole system clicked.
Because plain Markdown files can be read, written, and routed by an AI agent the same way a script works with any file. The entire knowledge base is just a directory of text files mounted in the container. When I send a WhatsApp message now, it lands in the right place in the KB without me opening Obsidian, choosing a folder, or writing a title.
The expression problem dissolved because I no longer had to “write.” I just had to think out loud. Two months to build the foundation. One night to complete it.
Why PARA
PARA is a simple organisational framework by Tiago Forte: four top-level categories — Projects, Areas, Resources, Archive.
Projects are things with a deadline and a goal. Active work. Areas are ongoing responsibilities with no end date. Things you maintain. Resources are reference material — topics you’re interested in, things you might need later. Archive is everything inactive. Out of the way, but not deleted.
What I liked about PARA was that it’s output-oriented. Notes don’t live in a topic category — they live where they’re useful. A note about microservices performance goes under the project that needs it, not under a “Microservices” tag that nobody searches.
The other thing I liked: it maps naturally to how work actually flows. Something starts as a project, becomes an area when it’s running steadily, feeds into resources when it’s something to learn from, and moves to archive when it’s done. The structure follows the lifecycle.
Why Obsidian
Obsidian is a local-first, Markdown-based knowledge tool. Notes are files on your machine. No database, no sync dependency, no lock-in.
The deciding factor wasn’t features. It was the file format.
Every note is a plain Markdown file. That means Lyra — my AI agent — can read, write, search, and route into the KB the same way any script can work with files. No API, no authentication, no rate limits. The entire knowledge base is just a directory of text files mounted in the container.
It also means the notes are portable. If Obsidian disappears tomorrow, the knowledge is still there. Sixteen years from now, a text file will still open.
Steph Ango, Obsidian’s co-founder, put it well:
“In the fullness of time, the files you create are more important than the tools you use to create them. Apps are ephemeral, but your files have a chance to last.”
That’s the bet. Not on Obsidian. On the file.
The graph view and backlinks are genuinely useful for discovering connections I’d otherwise miss. But they were a bonus, not the reason.
The Folder Map
The actual structure looks like this:
00_Inbox/Clippings/ ← web clips, articles, threads
01_Projects/ ← Kleos, Nyx, seevali.dev…
02_Areas/ ← Architecture, Career, AI, Personal Brand…
03_Resources/ ← .NET, AI/ML, Architecture Patterns…
08_Journal/YYYY/ ← daily journals, append-only
09_Showcase/ ← blog drafts, case studies, publishedThe numbers enforce ordering in the file system. 00_ through 09_ keeps the categories in the right sequence when sorted alphabetically. Small thing, but it matters when you’re navigating dozens of times a day.
The Decision in One Line
Plain files beat beautiful databases. Output-orientation beats topic-orientation. Ownership beats convenience.
PARA gave me the structure. Obsidian gave me the substrate. The combination made the whole pipeline possible — because a system that Lyra can read and write programmatically is a system that can feed content without manual intervention.
That’s what the next post is about.
Signal 3 of 7 in The Second Brain That Publishes Itself.