Learning a New Tech Stack with an LLM
While I mentioned a few weeks back that I'm now recovering from AI Psychosis, that doesn't mean I've given up on the technology completely. I'm writing more emails, and engaging with human beings as much as I can, and that part's been great. I still use agents as a glorified typewriter to generate code, and still very much get my hands in the dirt and sling code myself.
I've also gotten quite obsessed with the idea of writing backend services in Go, and bundling all of the frontend code and static assets in a single binary. I've been doing React and Angular for a long time, and have been pining over reaching for a new frontend framework. I eventually landed on Svelte / SvelteKit.
One problem though. I'm not fluid in Go, and I had never used Svelte before.
So I did a totally normal thing in the year 2026 and reached for an LLM. But instead of just letting the agent sling the code that I couldn't confidently review, I asked it to be my guide.
I could have used Claude's interactive learning mode, but opted to lay out the intention of the session and go from there. Simply put, I didn't want the robot writing code, just guiding me, and answering questions. I put heavy emphasis on sourcing live documentation and articles as references, so I could fact check a bit.
With the robot as my guide, I jockeyed the harness for the better part of a week. I started by building an app from the ground up, as I've done before in a number of languages. Having a solid engineering foundation, it was easy to push back on things, and easy to ask questions rooted in actual experience.
Similar to running an unattended agent, confidence has to be earned. In this scenario, I needed to build up my own confidence in what I was doing. That inflection point happened while writing a migrations script. After being guided on the "migrate up" portion, I decided I'd take the wheel and figure out how to handle "migrate down". Mostly successful, I turned the tables on myself, and had the robot review my code for a change.
I continued to write code, and fix it, and iterate the way I always have. We eventually hit the Svelte portion of the show, and the process restarted a bit, with me needing more guidance. By the end of the week, I felt confident enough that I could review generated code, and provide some semblance of feedback, and jump into the code myself if and when I need to.
This process was pretty solid, and quite similar to how I got up to speed on the PostHog Wizard and a few other codebases recently. Being able to jump-start coding without cracking the spine of a book, and having it taught against a real-world app is great.
Would I say this allowed me to become an expert fast? Absolutely not.
:wq
Like this drivel? There's a whole RSS feed of it, or subscribe via email.