Software as a Knowledge: When Code Becomes the Way You Think

I’ve barely opened Obsidian lately.

Nothing broke. I didn’t get bored with it. It took me a while to notice what had actually changed: the destination for my knowledge shifted. From text files to code. Or more precisely, to things an AI can execute.

When I Stored Everything in Obsidian

For years, anything I learned at work or while studying went straight into Obsidian. I linked notes to each other, tagged them carefully, and built up what I imagined as knowledge ready to use whenever I needed it. I was drawn to the Zettelkasten approach, and there were moments when I’d finish a note and feel a real sense of accomplishment — this one came out well.

Looking back honestly, though: only a fraction of those notes were ever actually referenced. “How to set up AWS IAM roles.” “How to run a retrospective.” Satisfying to write, but the next time the same problem came up, the cost of finding and re-reading the right note was surprisingly high. The knowledge existed. It just wasn’t being used. That gap bothered me for a long time.

What Changed When I Started Using Claude Code

After I started using Claude Code — a development tool that lets an AI operate directly on a codebase — I naturally began writing CLAUDE.md files, configuring MCP servers, and defining skills.

Then one day it clicked.

This is exactly what I was doing in Obsidian.

Take the principles I wrote in CLAUDE.md: “Simplicity first,” “Find the root cause.” These are almost word-for-word the same things I once kept in an Obsidian note titled something like “Work Best Practices.”

The one decisive difference: a text note is something I read. Software is something the AI executes. Whether knowledge gets used automatically or not — that gap turns out to be enormous.

Knowledge Crystallizing into Software

MCP Servers — Knowledge Externalized as Tools

I built an MCP server to feed GA4 and Search Console data to Claude for this site’s analysis. (MCP is the mechanism Claude uses to connect with external tools.)

Embedded inside that server is a set of judgment calls: which metrics actually matter, which queries are worth running, which signals to prioritize. In an earlier era I might have written an Obsidian page called “GA4 Analysis Best Practices.” But with an MCP server, that knowledge gets applied without any human in the loop.

Notes die if no one reads them. MCP just works by existing.

CLAUDE.md and Skills — My Thinking Process, Externalized

The CLAUDE.md configuration file holds project-specific guidelines and constraints for Claude Code. Globally, I maintain a rules/ directory: essentially a written record of how I reason through different situations.

This is precisely what “externalizing your thinking process” means. Before, I carried those judgment flows in my head, or buried them somewhere in Obsidian. Now they live in CLAUDE.md, and the AI consults them on its own.

Skills — the mechanism for teaching the AI recurring procedures — work the same way. “Run these checks before committing.” “Include this information when opening a PR.” Procedural knowledge that used to live in my head has become executable software.

The Gravity of Programming Languages Is Shifting

Traditional programming meant writing in syntax a computer could parse without ambiguity. One character wrong, and you got an error. The human bent to the machine’s rules.

Coding in the AI era feels different. “Simplicity first: try the simplest solution before reaching for a complex one” — written in CLAUDE.md — isn’t Python or TypeScript. But it is unambiguously code: the AI interprets it and acts on it.

Describing intent, constraints, and architecture in something close to natural language is becoming the core of programming. What I find interesting is that this might dissolve the boundary between knowledge workers and engineers. If you’ve ever written an Excel macro, built a Zapier workflow, or set up a Notion template, you’ve been doing essentially the same thing.

From “Writing” to “Making It Execute”

I haven’t abandoned Obsidian. It’s still where I draft thoughts and sort through ideas.

But how I handle knowledge has changed.

Anything I want to reuse or automate, I now convert into software. When a project wraps up, I find myself asking these questions almost automatically: “Can this decision be baked into an MCP server?” “Can this procedure become a skill?”

This isn’t only an engineering story. Excel macros, Zapier workflows, Notion templates — they’re all, at root, acts of converting knowledge into software. AI has dropped the barrier dramatically.

It’s less that knowledge is becoming software, and more that writing software has become the act of making knowledge. I wanted to put that feeling into words.