My order arrived 30 mins ago so things are a bit better after that plus some anime.
It's a ghost in a shell kinda day
12.03.2026 23:03
👍 1
🔁 0
💬 0
📌 0
My order arrived 30 mins ago so things are a bit better after that plus some anime.
It's a ghost in a shell kinda day
With coding agents, i spent about 2/3s of my time just telling them to debug things that not passing non functional requirements.
I think most of my claude skills are just different ways to build, test &. Profile the different projects.
Literally every train today has had the wrong announcements on it... The idea of the Cambridge to London train talking about places on the coast turned into a fun bit of trivia when it came to trying to find the train it thought we were on.
But on the underground, it's a bit more concerning.
Had my first case of claude code eating my project folder....
Definitely a surprise. Luckily I pushed everything a few hours ago so I only lost a bit of effort, but restoring things takes a while...
Rebuilding everything it's going to take ages...
Go go gadget CMake...
One of my learning projects at the moment has been to build a webserver from scratch, but starting from sockets and kqueue (I'm on a mac).
When I started using it, the number of unexpected cases and then actual edge cases exploded....
We really do use http for everything we shouldn't... 😅
Today can be summeised as follows
Grumble about trains
Chat about delivery
Opensource a bunch of stuff
Do battle with CMake
Grumble about trains some more
Later it's hopefully :
Chat about kafka
Chat about AI
And maybe even
Play card games
Here's a v0.1 of that #kafka sdk I mentioned :
github.com/Ugbot/zig-ka...
Goals are to be a drop in replacement for Librdkafka and be faster. I like @ziglang, it's got that c level clean feeling with 100% less CMake. Would have worked with the C code but felt like a change.
AI assistance was used
Observability is a chore.
It's definitely the thing you pay for so you don't need it, but it's really really annoying not to have it.
However it should not cost more than the solution you're observing!
Currently having a lot of fun with starrocks, and trying to get it wired up to be my logging and metrics store for my homelab.
Will fall back to clickhouse if I can't.
I really want to only have one platform for observability, as I don't want two storage setups for dev
Install and use UV. It's by far the cleanest option these days
There are times you can really tell what code the AI models are trained on and it really shows that they read a lot of the books I don't like.
I think half of what I tell the AI is "don't use mutexs", and another common one is: MVC is only for Web tools.
Today was another reminder of the importance of explicit over implicit communication in code, companies & life.
I've got an API surface that implicitly sorts the routes to hit the most specific ones first. But it didn't give explicit warnings or ordering to users.
AI DOESN'T GET IMPLICIT THINKING
Tfw you still have some edge cases to fix in your procedural locomotion system #gamedev
Expected top uses of claude code :
Writing boiler plate, checking specs, code gen.
Actual top uses of claude code :
Build scripts, fixing messy git repos, deal with virtual environments and containers.
It's actually removing the stuff that's just in the way of me writing code...
I got carried away when I finally got to play with Cpp corotines.
I built a Web framework from kqueue upwards.
Claude code helped with the typing, but mostly it seemed to attempt to use mutexs and synchronous code everywhere.
AI makes life easier but it doesn't remove thought.
Tis been quite the off couple of weeks, but Stargate sg1 is back on Netflix, so things are looking up😁
Using Librdkafka in @ziglang has been completely doable, but it's causing me enough issues that I want a cleaner lib...
Time to build a zig native kafka sdk.
Will open source when I get the last few APIs implementation (queues)
The fact that the new EU Digital Operational Resilience Act (#DORA)
Semantically overlaps with Devops Research and Assessment (#DORA) metrics is going to cause a lot of AI driven confusion.
The naming doesn't help😅
I know databricks has revenue, but their current funding drive is kinda crazy. I wonder who they are buying up next....
AI can do so many things, but it's ability to do ascii art on command line tools can not be under stated
I'm starting to use claude code differently as I run into the same issues again and again. The biggest one is file size bloat... It's totally fixable by hand but The AI can't deal with 6k lines of code...
I'm calling it a Smithers loop, as I'm giving it a different context each time we loop
Writing a sql query engine is a compiler. Except the language is mostly pre-defined, the users are mostly other tools and everyone thinks it's a solved problem (use postgres)
Today's challenge :
Teach an AI how to use kcache grind.
So far my AI Dev tools can use dtrace, valgrind and perf. But we need to go deeper....
Performance is one of the few ways to reduce TCO cleanly so performance is everyone's problem.
There's a new dresden files book.
It's one of those: "am I still reading it because it's good, or because I'm invested".
I enjoyed the earlier books at the time as a fun ride through a magical power fantasy, but on reflection I do wonder a bit.
Still bought it, but this is the last auto buy
I've been writing a database from scratch over the last few months with claude code and it's definitely an experience.
The most telling thing is that it really doesn't know anything you don't tell it.
It's a good faster keyboard but it's even remotely good at stuff with lots of moving parts
An enduring CS memory: a uni computer lab, ~2007.
I was failing at printf and realising how weird it is.
Nearby, students a year ahead were learning Java. I overhear: “wait… that’s printf.”
Turns out one of them had implemented ALL of printf in Java—rather than just learning how println works...
Claude code makes CMake only the worst thing in my day once a week rather than every day I work on a Cpp project.
It's weird what software tests can and can't catch. Especially with AI.
Unit coverage always feels like a vanity metric, as compilers and types should do most of that for you.
When it comes to feature level testing, you can end up catching all kinds of odd behaviour with randomised inputs
I'm all in on UV for this reason
Is this a test? It sounds like a test?
