It is VERY good. Although I went for the audiobook instead of paper or digital. It was an enjoyable listen, I think it fits it
06.03.2026 21:12
👍 1
🔁 0
💬 0
📌 1
It is VERY good. Although I went for the audiobook instead of paper or digital. It was an enjoyable listen, I think it fits it
I revisited this yesterday, wrote a more barebones version ("three regexes in a trench coat") and turns out it works almost as well, faster, and with a very simple codebase.
Garbell: skill and tooling for simple codebase exploration via progressive disclosure
github.com/rberenguel/g...
I'm also having fun encouraging Claude to use it and then asking "how was the experience? Anything we can improve?" I don't know what it is but this workflow makes me laugh
Latest newsletter - SVG is getting good out there, some good stuff in Gemini 3.1 Pro, my project to read in VR, a bunch of rando fun links, a handful of intriguing narrative gen finds. open.substack.com/pub/arnicas/...
I still want to see what edges of this design I can move around, it has barely 30 hours
Yep. The indexer breaks stuff around
It does automated chunking (but, a bit too naively), so should be able to handle relatively large (but not huge) documents. I'm still stress-testing it to see what needs improving, it's hard to create synthetic tests though. If you have a concrete idea I can throw to it, let me know!
I wanted to play around the RLM idea, but I really, really disliked the concept of having the context in Python variables. Instead of that (or my first idea of "put it in a sqlite db and make the agent write sql"), I wrote tooling to use as skills over a running daemon
github.com/rberenguel/c...
Sweet!
two rocky arches across a gorge connect to a central pillar and plateau upon which rests a single tree in front of a misty rocky background
made a ravine kind of scene in blender and ran it through my 160x100 8bpp qbasic engine. i kinda like it :3
#PixelArt #GameDev
If you enjoy this concept, you may enjoy:
- Murder, it wrote: an algorithmic whodunnit mostlymaths.net/murder-it-wr...
- Tanmateix: a timed logic reasoning game mostlymaths.net/tanmateix/
6/N, N=6
As usual with my personal projects, this has been half Gemini (idea refinement, initial planning and scope) and Claude Code (for execution), plus some more touches Gemini later to refine some areas. The icon is Nano Banana.
This is still kind of experimental, it will have bugs and glitches
5/N
Your goal is to find what service gives up first under that stress of the whole system.
The way it works internally, is that the game simulates 600 ticks (or however many are needed) until the first fault to appear. This is what you need to find.
4/N
WoMbat constructs a series of layered architecture (the depth is prefixed, as are the pieces that can be in each layer), plus an assortment of processing time, timeouts, memory per request, etc.
Then, it introduces a fault (say, a service suddenly taking longer to process). Your goal is…
3/N
After writing… wow, last week? Murder, It Wrote (and Tanmateix), both having different flavours of deterministically converting a scenario into a puzzle (CSP and Prolog, respectively), this weekend I thought "Could we have some sort of distributed system puzzle?"
The answer is _almost_ yes.
2/N
WoMbat: a distributed systems puzzle generator
mostlymaths.net/wombat/
1/N
Oh, on further inspection... I think I know the bug now. I changed accomplice clues and rules in one of my last minor updates, and I had a bad feeling. The count in the manifest is odd, it should have 3 people, and accomplice != innocent. I think there is an ambiguous double clue with accomplice now
Btw I'm ashamed that you precisely (as in, expert on "this"!) would have the worst possible first experience... I would say "give it another shot" but I'm scared now
It's kind of solvable, but the solution is undetermined by the clues. I still can't think how the pruner removed this clue. Aside from deterministic seeding I will add more logging (there is a log view hidden in the locations sidebar) so it's clear what clue is removed and why (as much as possible)
Hah! First I see that is completely not solvable, but indeed it is clearly not. I need to change all the randomization to be seeded so I can reproduce such cases. Accomplice not in the manifest is a known silly bug (forgot it).
I wonder what is the hole in the rule system in this one
Thanks :) The styling still needs work (I went wider instead of making a few look awesome). The cases turn out pretty well so far, always with a bit of a challenge. I need to come up with more fact types now :D
@arnicas.bsky.social this could be right up your alley (but, there is no generative AI except the ones that helped me with creating it 😬)
It can also generate a printable version (usually takes 2 A4, or a single A4 cut and folded in compact mode) in case you get a long case and would rather cozy it out in the sofa with a pen. I added this option today, it could have some issue I have not hit yet
It still needs a few UI tweaks (some of the styles are hard to read) but here's MURDER, IT WROTE
mostlymaths.net/murder-it-wr...
Close to infinite murder-style puzzles via a constraint solver. Except for some UI glitches (some fonts have a wrong color in some case) it works on phone and desktop
New (to me) coding agent use case: shrinking bug repros involving huge queries.
Essentially: "this query reproduces a bug: please binary search to shrink it."
Hey bsky I'm an artist and game developer transitioning away from twitter. If you like my art or my games help me out by sharing it around. There will be no shortage of RUIN related art for the foreseeable future.
#gamedev #pixelart
Excellent
Si que has comido
