Basil Berntsen

It’s got such a novel mix of promise and malevolence… like occasionally I feel that there is an emergent intelligence mocking me from behind the logs

I used to hate kubernetes out of ignorance. But now I'm happy to say that after learning kubernetes I now hate kubernetes out of knowledge

programming bootcamp in 2026:

“welcome to week 1. today we’ll learn how to open claude code”

week 2:
“prompt engineering fundamentals. please remember to say ‘please’ so the model doesn’t unionize.”

week 3:
“debugging. when the app crashes, we’ll paste the stack trace back to the ai and wait.”

LLMs will explain what’s going on if you ask. That’s become a critical part of my development loop: asking for explanations. I even have it in the persistent context markdown files that I am inexperienced with the specific tools being used.

On my main project that’s basically how I drive. Everything I’ve learned about the technical paradigm I’m working in I’ve learned well interacting with the Claude planner.

I did the reverse move recently, and I was able to organize such that both products use the same instructions more or less.

I use another model to help me review.

Everybody has their own tune when it comes to learning. Personally, I learn best by doing and reading, so I’ll give myself a task and every time I bump into a wall, I’ll unblock myself by reading something relevant. This has become much easier with AI now, but I used to search the internet.

Having pre-existing programmer skills and reflexes is absolutely valuable. Without them you’ll have to muddle through a bunch of unproductive work until you learn them. AI drives like a car in reverse until you constrain it with things like TDD.

I used to teach people coming from infrastructure backgrounds cloud. One of the things I would emphasize during the git lessons is that there’s always someone to ask questions and help you back out of accidents. That someone is now the AI. You can use it to learn anything you need to use it better!

MCPs are special- they let the text they generate pilot external systems. GitHub and Playwright are amazing examples.

My idea space expanded exponentially but my execution space only feels like it expanded linearly... the crazy part though is that I’m right at the beginning of using AI and I see myself getting better at it in parallel to its improvements.

It’s a place. It’s enough of a place that I’m willing to forgo the other places and their headaches.

More than their ability to extract money from companies, I am impressed by their ability to extract money from people at home. I pay them now, and I’m considering paying them even more because the more I pay them the faster my side project gets built.

I’ve spent my whole life as a function before form kind of guy, but find myself now trying to build a sudoku app that will be visually appealing. I’m hoping AI can help me compensate for my miserable sense of colour and aesthetics 😂

I was able to set options not exposed in the vs code addon using a prompt- but rewinding works well for me out of the box.

They’re still there, you just have to ask Claude to help you edit them into the correct settings files.

I’ve been trying Claude Code after months with GitHub Copilot and it is amazing. GitHub Copilot was amazing too though- I made excellent use of agent tasks and the vscode integration

I’ve only tried the vscode plugin and that’s a bit buggy but usable

It’s a hard dependency for some common technologies in that space, off the top of my head Ansible.

I had never looked into the GIS stuff either and I had no idea you could do things like query by distance without an external tool

Now I reach for the right tool for the job, instead of the best tool I know

YouTube video by Awesome Postgres can replace your entire stack...

I have not paid enough attention to #Postgres- m.youtube.com/watch?v=1qs9... was eye-opening

I didn’t need it but I suppose it depends on what you’re doing. I work on too many different code bases to not use compartmentalization. The way I do that is dev containers in VS code, and if I’m on Windows that means WSL and docker for desktop.

I don’t understand people who don’t keep themselves in the loop. I tried at first because I underestimated the complexity of what I wanted and overestimated the ability of AI to stay on target. Now every time I release too much latitude to AI, I regret it.

Containers can protect your local environment but they can’t protect something that’s being accessed over an MCP. If the permissions include overwrites, I think the model can choose whether to ask you or not.

Now every time an agent builds something, it runs a linter with some simple rules that can hint there might be smelly code, and then has explicit instructions to make judgements based on the actual smelliness and either fix the smell or comment out the linter warning.

This is outside of my normal experience so I just kind of blindly started building. When I started to get results that didn’t match what I wanted, I dug and discovered opportunities to improve the quality of the code. Given my lack of experience with these languages, this was all with AI.

So I’ve been surprised at how good certain models are at the high-level stuff. I just kind of blindly stumbled forward until I realized I produced an application that works but resists new features because of a bunch of smelly code. I learned about this and how to address it using one of the models.

Cloud is a way of hosting things. It’s kind of plumbing for the interesting stuff. I’m responding more from the perspective of my side project which is AI engineering a sudoku app using languages I barely know.