Weeknotes 268

RubyConf 2026 Recap

There is a new ‘killer app’ for Ruby

I’ve been hearing that about Hanami and friends for a long time, but no one I know has used it or barely even tried it. I wish it were true.

How we tracked down a 16-year-old SQLite bug – Lord, let me be so successful that I encounter a bug such as this.

And can I take this opportunity to congratulate Alex Chan on the sub heading “The writing on the WAL” 👏

The anti-AI contingent amongst you might be satisfied to know that this week was not a good experience for me. The roller coaster is currently on the descent.

I am on a journey to find the line between “watch it’s every move” and full-on “what is computer code anyway”, and I’ve not found it yet. I don’t want to give up control, but watching the LLM’s every move feels constraining. I often feel that I am the one in the way of getting things done – as has so often has been the case in the before times. So I’ve been trying to get out of the way.

Getting started is one of my biggest struggles as a builder of things. I find the enormity of the thing completely overwhelming and paralysing (feel free to apply your own diagnosis here). Common advice for scenarios like this are to break big things down into smaller things that can be more easily worked upon. No disagreement there. The adviser never acknowledges that that in itself is a task too great for many to complete, including me. This has been a frustration of a lifetime.

In order to get out of the way and break things down into chunks that can be worked on I’ve been trying out Matt Pocock’s skills. They are seemingly very popular and the ideas within resonated with me. Of course, skills are mostly just text someone else wrote, but as a standardised place to start from I think they make sense. I made a concious effort to try them out this week, in particular, the /wayfinder skill.

wayfinder takes an effort too big for one agent session — an idea whose destination you can name but whose route you cannot yet see — and charts it as a shared map of decision tickets on your issue tracker, then resolves them one at a time until the way is clear.

You tell it what you want to build or have as the output of the session and it basically asks you a shit ton of questions, and then a shit load more, until you have something actionable.

I found that it asks you lots of really interesting questions about things you had never really even considered before. This is useful. Programmers, in my experience, are extremely predisposed to just writing code immediately upon hearing that something needs building. The way these questions make you think about the system are a good thing.

I turned the output of this interrogation into a spec, and then tickets which could be implemented. However, when it came to implementation I found that the result was often massively over-engineered. It would take a long time, churn, churn, churn, and then show me a complicated solution to a security problem I didn’t know I had. This had me asking myself if I was opening up my system to security issues or simply course correcting the LLM into not over-building.

This week I’m left confused and without much progress. I think that the questioning is very useful, but I don’t want to specify the whole system upfront and I think maybe that’s what ended up happening. I need to be far more precise in telling it what I want out of it.

Run AI agents safely in local sandboxes – Another place to run agents.

With all the homelabbing I’ve been doing over recent months my docker-compose.yml had gotten a bit messy. I wanted to clean it up but before that I thought some automated checks were in order, so I’ve now got dclint, yamllint and kics setup via a GitHub Workflow running on new PRs. Whilst I was at it, I added hadolint for Dockerfiles too despite only having a single custom Dockerfile in use at the moment, but it doesn’t hurt.

That’s a lotta lints.

dclint in particular has been very useful in helping me normalize a lot of the config. The docker-compose.yml is now in a much better shape.