The Invisible Software Stack
AI is changing our personal software stack to be hyper personal
I’ve switched back to Obsidian because of AI, but not for the reason you may think.
Obsidian hasn’t developed more powerful AI features.
Obsidian hasn’t added any killer features that I was missing.
Obsidian’s plugin ecosystem hasn’t gotten any better.
The killer feature - extensibility - has been there since the beginning, but AI coding agents have unlocked access.
I can run Claude Code, Gemini, Codex, and any other agent I want on top of my entire Obsidian vault.
I can use Claude Code to author Obsidian plugins to do whatever I want.
I can make Obsidian the central nervous system of my workflow - hyper-personal and fully automated
I can now make this platform do exactly what I need it to do, and I don’t have to worry about whether it works for anyone else.
AI coding agents have changed my relationship with the software. I no longer look for the most complete or polished application. I look for the one that will give me maximum ability to shape my own experience by building inside of it.
When I can’t find a tool with that flexibility, I build it myself. Agency and control are the new feature foundations.
Designing My Own Workspace
In the last few weeks I’ve built a handful of applications I needed:
A call transcriber that listens to system audio, performs real-time transcriptions, and lets me chat with those transcriptions (Yes, just like Granola)
An Obsidian plugin to sync Zoom transcriptions directly into my vault
A tool that inventories my test suite and helps me understand and refine it
A worktree management tool that launches containers for isolated dev environments (yep, there are tools, I wanted one that didn’t exist)
None of these are particularly novel ideas. Transcription tools exist. Ways to export Zoom data exist. Test management tools exist.
What has changed is that, in the time it would take me to find an imperfect tool for any of these things, I can often build a perfect one. All that searching, testing, potentially paying for a tool I don’t use… all of that goes away. Not for every use case, you still have to make a build vs. buy judgement - but I’m finding for many situations, the tool I can build is preferred.
And I have full control over it. That’s the part that matters.
Control Over Compromise
The call transcription tool is a good example. I’ve been using Granola for a while now and I really like it, but my transcripts are locked away inside Granola and I have limited control over what I can do with them. I can pull them into Obsidian or Notion, but that’s not actually where I always want them - it depends on the context. Granola also insists that people in this world only have one context for calls - business - a fundamental assumption that doesn’t apply to me.
I wanted to control what happened to the transcription after a call ended. I wanted to control the quality of the transcription. I wanted to be able to take action on the transcription during the call - something I could never do with a tool like Granola until they built that feature in. However, in an afternoon, I had a functional transcription tool that ran in my terminal and could be extended easily.
I might be the exception here, but I am not super keen on waiting for the rest of the world to figure out how I should be leveraging AI in my life and tools. I’d like to be exploring now. In order to do this exploration, I need to be able to use AI to extend the tools I’m using.
The Economics are Shifting
I have a complex application with a flaky test suite. I got tired of fighting with it.
So I built a tool to inventory my entire test suite. It runs each test and tells me how long it takes. It explains why each test exists and what it covers. It walks through the flow of the application that each test reproduces. I can ask it for whatever I want, and then tune my test suite based on that information.
Here’s the thing though, I have no idea if a tool like this already exists. I didn’t bother looking.
That’s what is changing. The economics have flipped. If something seems like it’ll require searching, evaluating, and potentially paying for, I’m just going to try to build it first. If I have a hard time building it, then maybe I’ll look for commercial software. That’s the opposite of how this would have worked even a year ago.
The Application Inside the Application
For people who are already working in codebases, there’s another layer to this. Within my more complex applications, I’m building sub-applications:
A tool that shows me the design system components and color palette available
The test inventory I mentioned earlier
A CLI built specifically for AI agents to interact with my Linear ticketing system
I can build control surfaces directly into the repository I’m working in as tools specific to that codebase, without integrating external solutions. Maybe that’s wasteful if there’s a great tool I’m not using. But I’ve been burned more times than I’ve been saved by pulling in outside tools, it’s just not hard to build my own.
Over time and at scale, these tools become more commodity, I know that, and using your own special thing becomes a liability. But I’m operating mostly solo here, so personal friction reduction is most important to me right now.
The Invisible Software Stack
Here’s what I think some people are missing about AI coding agents: there’s an entire class of software being created right now that nobody sees.
This isn’t new in the development community. Developers have always built personal tools. But now it’s spreading to product managers, to people who traditionally would be less technical. They’re building up inventories of software they’ve made for themselves. Maybe some of it gets shared, but most of it won’t work for others because it was built for a very personal purpose.
This software is invisible because it’s not released. It’s not on Product Hunt. It’s not in app stores. It solves one person’s problem in one person’s context, and that’s enough.
A Different Way of Looking
AI coding tools aren’t just about building apps faster or replacing developers. They’re about agency. The ability to control your own tools instead of waiting for someone else to build what you need. The ability to fix what’s broken instead of working around it. The ability to approach problems you would have ignored because the solution didn’t exist and building it yourself wasn’t realistic.
Right now, this probably works best for relatively simple applications. But the line of what’s “simple” keeps moving. And the fundamental shift: that you can be the one who solves your own problems, well, that’s already here.
What are you building?
Have you solved something for yourself and only yourself? Have you built an app inside your coding project?
I wanted to share my own recent experience with using AI in website development; though it is not on the same level of sophistication as your's, Aaron!
I like helping non-profits with their projects, which I find more fulfilling than chasing projects for profit. A recent one is the Fort Collins Community Opinion Platform, a grassroots effort to bring back something like a 'Letters To the Editor' columns, which were once prevalent in newspapers, but to do it online.
My hosting platform is Siteground, and since they had been touting this new whizbang AI tool of their's, called Coderick AI, I decided to give it a try...& here are my conclusions at this point:
After submitting a short paragraph describing the website's basic functionality, it was somewhat of a revelation to see that this tool did indeed produce a bare bones website -- in all of five minutes!
However, although the code appears to all be in Javascript (using Chrome Develop tools), it is utterly unmaintable. The only way to change the website is to regenerate it. After conferring with Siteground tech support, they confirmed there is not really a code base that can be used, nor any dashboard by which to modify the website appearance.
The bare bones website is just that: spare and functional. Not example something you'd be proud to show your client! However, if I can get the tool to add some attractive styling...or import some pictures...it might work.
What I proved to myself is that using AI to build websites is certainly possible; but not terribly useful at this point...unless I make some breakthrough is how the site is generated.
Any suggestions appreciated!