Language, not pixels

Building software is a lot of fun right now.

I keep seeing people complain that we're producing a lot of software, and that a lot of it is garbage. I get it. But I don't think this is new, people were building garbage software long before AI existed. We just do more of it now, because it got cheaper to do. And there's nothing wrong with that.

I also don't think all software has to be broadly used. It's totally fine to build something for yourself, put it on GitHub, and share it with the world. If some people end up using it, great. If nobody does, that's also fine. Who cares. You built something and you put it out there, you created something for the world.

That's how I've been feeling about AI lately. It's becoming more and more a means to an end.

Trading code for a black box

I've always been into building things. A product kind of person. I still look at code sometimes. But lately I've been putting myself in a different way of working, where I kind of trade the code away. What if I don't look at it? What if it's a black box? What if I focus on the user experience instead, experience the app the way the person using it will, and steer the agent toward how I think that experience should look?

It took me a while to get comfortable with it. I still run periodic sessions to check the code and keep the architecture heading in a good direction. But day to day I'm much more on the experience side, and honestly it feels natural.

It started with me wanting to own my data

This whole "everything goes headless" thing started while I was managing a company. I noticed that a lot of the operations we had to run depended on interfacing with the UI of some SaaS. I wanted to automate that. I wanted to call an API, or have Claude or Codex call an API for me, and I felt completely constrained. Why?

So I started moving away from those platforms. CRM, sales, finance. I started owning the data, and by owning the data I control how it's accessed. Sure, I could build a dashboard for it. But I realized I don't really need a dashboard. What I need are interfaces for my agents to access the data and do something with it.

That opened the question: couldn't everything be headless? If I have my assistant on some bot that can reach the data, what is really the value of these platforms? I wrote about headless development a while ago, focused on the dev loop. This is the same idea, applied to everything else.

This post exists because of it

The first place that materialized was my blog.

I built a tiny blog engine, a simple version of a CMS where I push a blog post and it gets published, made accessible to the internet through a blog front. But everything else, the administration, is done through an MCP interface connected to Hermes, which I talk to through Telegram. Right now I'm dictating this whole post as a voice message on Telegram, and it gets committed and pushed straight to Gesttalt. I wrote about the publishing pipeline here.

It's fun because it's just me. My ideas flow from the microphone of my phone, we iterate a bit on the grammar together, and I push it to my site. That's it. My ideas, my content, nothing in between. Just a different way of interfacing with technology.

I took the same approach to social platforms. Same story. I skip the writing entirely and go straight to Telegram. "I have this idea, let's post something about it."

The vault, headless

I used Obsidian for years. It never quite clicked.

I understood the idea of semantically connecting notes. But me being the one making those connections felt wrong. The bigger the graph gets, the harder it is to connect everything by hand. Once I had tools that can do that semantic analysis and understand the space between pieces of content, connecting them became a task I didn't need to carry anymore. I wrote about the graph idea back in 2020, and I wanted the graph to be an implementation detail. Now it finally is.

I buy the Obsidian idea that you own your vault, that it's offline, that it's yours. But if I self-host my notes and I'm in charge of them, I should be able to have multiple editing interfaces, one of them being just Telegram. Obsidian can't really give me that, and depending on their syncing engine felt wrong.

So I took the storing part of Obsidian, pushed it to a server, added a semantic layer on top, and then added MCP and REST interfaces. Now I can paste a document I found on the internet, with some cool ideas in it, straight into Telegram, and it gets persisted with the right semantic information so I can retrieve it later. Or I can just ask, "find that document I saved a month ago", and it brings it back so I can read it.

The whole thing becomes super headless. And that's amazing.

A trainer in your pocket

I was going to the gym with a personal trainer. I stopped, because it got a bit expensive and I felt I'd learned everything I needed to continue by myself. But I still wanted a system to iterate on my workout template, track my progress, and have a conversation with it. "This isn't working." Or, in my case, "I have an injury, how do we make sure we're careful with the hamstrings? Is this exercise safe?"

So I built that too. I call it Ted, after Ted Lasso. It's on GitHub, I run my own instance, it connects to Hermes, and my trainer now lives on Telegram.

So on Telegram I can be posting an idea to my blog, chatting with my trainer, looking up something in my notes, checking my finance data. It goes on and on.

Text, not pixels

Language is one of the most primitive tools we have. But when technology showed up, we went straight to UI, pixels, buttons, forms, pages. Which is not really natural.

I'd love to see everything gravitate toward text instead. That's how I think things should be.

What this means for software companies

This poses an interesting challenge for a lot of businesses, because they invested years in their graphical interface, and in ecosystems of plugins and extensions built by developers. There's an incentive for developers to build on those ecosystems, and the platform markets itself by saying "we're better because we have all these plugins."

If we move to a world where, first, you can customize your software, you don't really need a marketplace, because any extension you need you can build yourself. And second, if you interface with the tool through text, you don't really need the UI at all. You could be talking directly to a database, plus a thin layer of business logic on top.

The more you think about it, the more you realize a lot of these businesses will probably stay afloat for a long time. But they become really simple in terms of the value they provide. A lot of that value was always perceived, created through marketing and other tools, more than it was real.

What it means for me

For me and for Tuist, that means focusing on infrastructure. Delivering value through infrastructure, because I think the value provided by software is going to collapse, and become very thin on top of some messaging application.

It's been a lot of fun to watch, and to build in that direction.

The whole loop, closed

I recorded all of this as a ten minute voice message on my phone. Now Hermes and I will iterate on it in Telegram until the title feels right, I'll push it, and tell it to announce the post on X and Mastodon. That's the whole thing. Just me, my ideas, a microphone, and a couple of agents.

Building software is a lot of fun right now.