Barry O'Sullivan

I've been playing around with Claude AI by building a boardgame app. Claude is surprisingly useful if you know exactly what you want it to do and understand clean system design, e.g. DDD concepts, separation of concerns, testing... You can't trust the output, but as an assistant it's solid.

Your data model is your destiny Your product's core abstractions determine whether new features compound into a moat or just add to a feature list. Here's how to get it right.

The abstractions you chose at the beginning will come to define your product. Great read on the value of design.

open.substack.com/pub/matttbro...

twitter is currently a frankenstein combination of 4chan and tumblr. we all know this. but bluesky is a different beast that i don't think people born after ~2004 fully grasp. bluesky is a frankenstein combination of tumblr and somethingawful

Is this the status quo of legacy? This is very selective and assumes your legacy is what should be expected, that's not the reality of testing in legacy.

(Uncut) Metal Gear Solid V: The Phantom Pain & Ground Zeroes - YouTube

If you want a fun, and ridiculous palate cleanser, I feel recommend @chipcheezum.bsky.social and @mynameisgrant.bsky.social doing a playthrough of MGSV. Honestly one of the best letsplays going.

youtube.com/playlist?lis...

You are now free.

The AI coding trap | Chris Loy If you ever watch someone “coding”, you might see them spending far more time staring into space than typing on their keyboard.

chrisloy.dev/post/2025/09...

Claude told me that it doesn't automatically load CLAUDE.md files, despite the docs saying so.

When I asked why it contradicted its own docs, it went and scanned them (consuming loads of tokens) and told me I was correct and it was wrong.

You really can't trust this thing.

Told it to use the laravel sail docker image instead, that failed too. Finally told it to just use laravel sail instead. That looked like it worked right up until I got the default apache page and not my app.

Pretty much. Asked it to create a laravel default laravel app, which worked fine. Then asked it to setup a dockerfile that could boot and run said application. This is where it fell down. It chose an outdated image initially, which failed to get deps and failed.

I tried using claude to setup a simple laravel app with docker compose and it kept messing up. Docker was misconfigured repeatedly. Once I got it working the webserver wasn't even pointing at the right dir.

Absolute waste of time. I'm gonna do it manually instead.

The Devourer

"Searching for my rhythm in a maze of algorithm" is such a banging lyric.

open.spotify.com/track/6u3sO5...

Thanks for that Padraig, really appreciate it.

ICYMI: I wrote about how your domains and bounded contexts don't map 1 on 1. #DDDesign
verraes.net/2025/08/doma...

While I'm willing to believe that AI can help with modernization of legacy, it's not going to fix those poor design choices one made along the way, it may even erode what calcified over time - there's probably no test suite worth speaking of that is going to tell you that happened.

How cryptoic.

This entire thread fills me with joy and horror.

No joke, an advert for a device you wear on your arm that "blocks EMF", had a picture of a dude wearing a tinfoil hat.

We are terrible at refactoring Software engineering, as a profession, is terrible at refactoring. A hot-take I know, but I think I can convince you. After writing about [skill acquisition and up-skilling](https://barryosull.com/blo...

We are terrible at refactoring. Many of us lack the skill and we are not incentivised to gain it, thus our technical systems rot over time.

What can we do about this?

(New blog post)

barryosull.com/blog/we-are-...

I just saw an advert on Reddit for a digital tinfoil hat . . . what kind of profile have they created about me?

for all people mock it wikipedia is genuinely one of the wonders of the modern world

The end justifies the means. But what if there never is an end? All we have is means.

Why hasn't AI improved productivity? Simple, LLMs are good at generating reasonable looking language, but that's never been the bottleneck. Infact, they actually make things worse, as they flood our systems with noise, putting more pressure on the bottleneck.

A mimic bed, now there's an idea.

This was my experience with a take-home exercise. We had a 30min chat about my implementation and it was clear to them that I wrote it as I could answer every question. Impossible to do with something you didn't write. The interview is essential and makes take-home exercises worthwhile.

if you ask it for a thing that looks like an answer, it gives you a thing that looks like an answer

if you ask it for a thing that looks like a source, it gives you thing that looks like a source

if you tell it that it's wrong and ask it for a thing that looks like an apology, that's what you get

LOLs per Employee: a vital metric.

Shenanigans are a sign the team has sufficient energy in reserve, signaling readiness for handling emergencies and unplanned work. Play usually reflects underlying trust, cohesion, and resilience.

Encourage it.

"Vibe Code Cleanup Specialist" is being added to a lot of LinkedIn profiles. Looking at the bios, rarely see any mention of refactoring, legacy code, tests or separation of concerns.

I suspect "Cleaned-Up Vibe Code Cleanup Specialist" is coming next.

You will never be able to make "data driven" decisions until you realize just how much your decisions are driven by emotions: fear of missing out, defensiveness of your favorite idea, apprehension over something unusual.

That's because there's no such thing as "the data" - only interpretations.

We've been putting some much-needed love into our virtual DDD community! We're evolving beyond just Domain-Driven Design to become the worldwide online community for designing software to solve complex business problems, organised in a decentralised way.
-->