Personalized software really is coming, but not today. Maybe tomorrow?

Matt Sayar

When I posted about building a custom math game in ten minutes, one of my favorite comments was

This is so awesome. Personalized software really is coming.

Recently, Anthropic released their version of an agent that can integrate with your codebase and make changes for you. The name, Claude Code GitHub Actions, rolls right off your tongue. I still don't know where they announced this; I just found out about it via a random comment on Hacker News, and I can't find a way to subscribe to Anthropic's newsletter. 

I wondered if this could be the agent that helps achieve that dream of "personalized software." I've been keeping a scratchpad of feature requests for Publii, the tool I use to build my site.

I originally intended to write this code myself and submit the PRs to the Publii GitHub, but open source maintainers are getting overwhelmed with shitty vibe-coded PRs and are actively rejecting them.

But why not fork and create my own? First things first, I need to figure out how to build locally, then go for it.

Side note, why is building locally always a pain? You try to follow the instructions, you get errors every step of the way, Google/GPT the errors, update/switch versions of tools and packages, slowly chip away at it until it finally builds... Anyway, I eventually figured it out, with very low confidence in its stability.

webpack 5.99.9 compiled with 431 warnings in 4478 ms

Programmers always ignore warnings. I'm sure it's fine... (spoiler: it isn't fine)

Unleashing Claude Code on my personal build

OK, let's try this out. I installed Claude Code to my fork, and created a simple feature request. I started with something simple: let people click the "Share It" button at the bottom of this page to share with BlueSky. Should be easy, right? Small scope, lots of similar functionality already built in...

It was as easy as asking @claude implement this feature based on the issue description. In only six minutes, it churned through it and created a branch capable of merging in just a couple steps.

Pretty cool! Have I achieved the dream of ✨personalized software?✨

No.

The first issue is that every time I build a new version of my own Publii, it opens as if it's a first-time install. That means restoring my site from backup, reconfiguring my site's server settings, re-enabling dark mode, etc. I'm sure there's a way to "publish" my own local copy that's permanent and not ephemeral, but I have so many build errors (see above) that I don't trust myself. I understand this is a Me Problem.

The second issue is that the code doesn't work. After rebuilding everything, the BlueSky link still doesn't show up in the config settings. Just looking at the code, it looks like it should, but it doesn't. 

It's supposed to be right here!

After spending a couple hours trying to get my local copy to build (and yes, I double-checked that I was using the right directories etc.) I wasn't interested in debugging this anymore. Especially after I saw what the Claude-created SVG of the BlueSky logo looks like:

That ain't it 💩

With more patience, I'm sure I could successfully build the Publii app locally. I could instruct Claude to make a better BlueSky logo with some examples, and I could figure out why the config setting isn't showing up. Then I could publish my own local Publii app and live the dream of personalized software. But I don't want to maintain my fork with the latest official build, I don't want to figure out why the ls-all npm package isn't installing right, I don't want to debug AI code, I don't want to write a build script, I don't want to do this all over again for my next feature request, and I don't even want to try a moderately-complex feature request at this point.

I'm pretty bullish on LLMs and how they help the development process. If this was my job, I wouldn't be writing this, and instead I'd be figuring out how to make this work. If there's a more critical feature I want, I would also try harder. It's all very promising! But it's going to be a while until personalized software is accessible for everybody. It's likely closer than I think.