30 Days of Claude Code + Obsidian

What Actually Stuck After a Month

I’ve been running hands-on Claude + Obsidian workshops. Seeing what actually works, and where people get stuck, shaped the last 30 days.

A Warm Coffin

Most knowledge systems don’t fail because they’re bad.
They fail because they’re comfortable.

They work just well enough that you never feel an urgent need to change them, even as they slowly stop helping you learn, think, or adapt. There’s no dramatic breaking point. Just gradual stagnation.

That’s what I mean by a warm coffin.
Something comfortable enough to stay in while it quietly restricts you.

That’s also why people defend bad systems so aggressively. Not because they’re good, but because they’re familiar.

Knowledge management evolved roughly like this:

oral stories → writing → computers → folders → search → databases → networks

Each step was a genuine leap forward. Each one also reframed the previous approach as a polite historical artefact.

Computers gave us scale.
Folders gave us containment.
Search gave us hope.
Databases gave us structure.
Networks give us relationships.

Obsidian lives in that last shift.

People still stuck at the folder, search, or database stage are essentially cave painting, just with better fonts and recurring billing. Obsidian already escapes that trap by treating knowledge as a network instead of a filing system.

But that’s a low bar.

Compared to how people will be using Obsidian once AI is properly integrated, you might already be falling behind. You might feel technically savvy compared to your parents. Compared to people actively using AI to learn better, you’re still operating inside familiar limits.

What’s different now isn’t the tools themselves. It’s the rate at which they evolve.

We’ve seen transformative tools before. But they spread slowly enough that falling behind was temporary.
Today, the pace of change is accelerating.
When change compounds, falling behind gets harder to recover from.

The silver lining is that these are also the most powerful tools we’ve ever had for learning faster.

The irony is that the faster you learn these tools, the faster you can learn everything else. The advantage compounds.

That’s why the most important skill over the next five years isn’t a specific technology. It’s learning how to learn, and these tools finally make that practical instead of aspirational.

And uncomfortably, the only real defence against AI isn’t avoiding it.
It’s wielding it properly.

I wrote about my first 48 hours with Claude Code. Then my first 14 days. Now it’s been a month.

I wish I could say the novelty is gone and the honeymoon phase is over. But I’m still the guy cornering strangers at dinner parties to explain how fucking sick my Claude + Obsidian system is.

What stuck

Daily logging without the admin hell

Everyone loves talking about “building a second brain.” Nobody loves manually creating notes, linking people, tagging projects, formatting timestamps, and choosing the correct emoji.

That’s not thinking.
That’s data entry.
And I refuse to do data entry in 2026.

Claude does it now

I say, “logged meeting with Dan at 9:30,” and Claude handles the timestamp, tags, emojis, and links. Even when the actual filename is “Dan Blonde stocky pierced nipples marathon gym sauna guy.” It matches by first name and carries on with the level of professionalism and lack of eyebrow-raising I would expect from a loyal personal assistant.

Unlike some of you.

What used to take minutes now takes seconds. Capture stopped being optional. My vault went from patchy to embarrassingly detailed. My future biographer is going to need a lot of hugs.

The most valuable thing Claude does isn’t clever.
It’s clerical.

What evolved

From capturing to producing

This surprised me.

At first, Claude was purely for input. Logging. Formatting. Organising. Then that became normal. With speech-to-text dictation (thanks, Wispr Flow) feeding straight into my Claude personal assistant, capturing felt effortless.

Then it clicked.

When I sit down to work now, I don’t start with a blank page. I start a work session and ask Claude to surface everything I’ve already thought about the topic. Relevant notes. Previous work sessions. Old arguments. Half-formed ideas I forgot I had. All of it appears directly in the note I’m working on.

Instead of asking Claude to generate ideas, I ask it to resurface mine. This article came together that way. Nothing here was invented. The raw material already existed.

Writing feels easier because the hardest part was never writing. It was locating your own thinking. Now I start from my past work, decide what matters, and assemble the argument.

I’m not starting from other people’s information.
I’m starting from mine.

The same applies to sources

Instead of letting articles, videos, and threads pile up and radiate low-grade guilt, I process them as they come in. Key points are extracted, tagged, and stored.

“I’ll read this later” is a lie.
“This is already processed” actually ships work.

The thinking is mine. The organisation is automated. When I sit down to produce something, I start from my material, not a blank page and not something generated for me.

I’m building on my own knowledge, not outsourcing my brain.

Speed via a vault graph

I asked Claude to generate a simple JSON graph of my vault to speed up note lookups. It’s essentially an index. Instead of searching through files every time, Claude checks the graph first.

This came from a comment on the previous article, and it turned out to be a clear improvement over the default “go search the vault” approach. Faster lookups. Fewer broken trains of thought.

Security concerns didn’t magically disappear

One thing that hasn’t faded with time is my discomfort around security.

Using AI like this means sending personal context to a large corporation. Not abstract data. Your notes. Your work. Your thinking. That’s a real concern, and it should be treated as one.

Most setups today involve some level of external processing. That’s the tradeoff. There are people running fully local or hybrid systems where sensitive data never leaves their machine. I haven’t settled on my own version of that yet, but I’m actively looking into it.

For now, it’s a conscious tradeoff. You’re trading some privacy risk for very real upside in leverage, speed, and learning. That doesn’t make it reckless. It does mean you should understand what you’re opting into.

The point isn’t to ignore the risk.
It’s to be aware of it and decide whether the reward-to-risk ratio makes sense for you.

For me, so far, it has.

What I dropped

One massive CLAUDE.md file

Long instruction files fail for the same reason terms and conditions exist: they’re technically correct and functionally ignored.

Claude behaves exactly like a human here, just with a slightly better memory. The longer the instruction file, the more slips through the cracks. Past a certain length, you’re not giving guidance. You’re just yelling into the void.

The fix was to compartmentalise.

Specific Claude Skills for specific actions. Small, focused instruction files loaded only when needed. Claude follows these. The monolith never stood a chance.

The four I use constantly:

Daily Logging
Handles timestamps, tags, emojis, and links. Used constantly. Losing this would hurt emotionally.

Process Meeting
Turns raw transcripts into summaries, action items, and follow-up emails. An 80-minute call becomes something usable in minutes.

Process Source
Extracts summaries and key points from articles, videos, and threads. My source folder went from graveyard to tool.

Claude feels like a very fast intern who never complains and never questions why I saved a hobby horsing video tagged #bachelor-party-idea.

Week Review
Reads all seven daily notes, summarises the week across life areas, and selects three photos prioritising people, experiences, and milestones. What I used to skip now takes 90 seconds and actually happens.

What 30 days taught me

The compound effect is real.

When logging takes no effort, you log everything. When reviews are automatic, you actually review. When sources are pre-digested, you use them. When connections take seconds, your vault stops being a graveyard and starts being useful.

The biggest shift wasn’t any single feature. It was the collapse of the gap between thinking something and having it captured, connected, and retrievable. My vault went from storage to a thinking aid. That’s what everyone promises. Almost nobody delivers.

I’ve used Obsidian for three years. Capturing was never the problem. Producing was. Having your own processed material changes that completely.

If you want to learn this setup properly

I’m running small, hands-on workshops for people who haven’t touched Claude Code yet, or who want to move past “I installed it” into actually using it.

We set everything up on your device, and immediately work through real use cases:

• Daily log automation

• Project and meeting creation

• Pulling web content straight into your Obsidian templates

If this article made something click, this is the fastest way to make it stick.

Sign up here

Much love,
Dee
Construct by Dee

Website | YouTube | X / Twitter

---