site banner

Tinker Tuesday for July 21st, 2026

This thread is for anyone working on personal projects to share their progress, and hold themselves somewhat accountable to a group of peers.

Post your project, your progress from last week, and what you hope to accomplish this week.

If you want to be pinged with a reminder asking about your project, let me know, and I'll harass you each week until you cancel the service.

2
Jump in the discussion.

No email address required.

I changed to monster-projectile collision model from circle-circle to circle-rectangle, it should give me a bit more flexibility with what I can do with the projectiles. I still have to add an offset factor so the collision box isn't necessarily centered at the projectile location, and I still have to test it fully, but so far it seems to be doing what I want.

During the tests I was playing with the parameters the simulation. I managed to smooth out the monster-monster collision so they don't get stuck on each other so much. I also thought I could later incorporate some of the parameters into the game mechanics, like adding a time slowdown, or a "gravity bomb" that could pull all the monsters and crush them, but for now that's heading to the backlog.

How have you been doing @Southkraut?

Tinkered up a little globe view, colored the thing by biomes (latitudes really), set it to spin on its polar axis, put a fixed sun in the sky and a little moon into orbit...nothing interactive, just a little backdrop.

The bulk of what I did this week was UI work. Got to implementing a main menu, and a strategic management UI. Made good progress, until I realized today that I'm using the wrong framework for what I'm trying to do. Using Slate instead of Widgets. I started out with Slate back when I tried to do as much as possible in C++ vs in the Unreal Editor, but right now I'm actually leaning in favor of taking the WYSIWYG approach to UI design; so long as I retain control over it via code. So right now I'm trying to get it all set up as Widgets instead - a slow process, doubly so since AI can't do that for me.

Speaking of AI, I've been keeping it on a shorter leash, letting it work in smaller chunks and reviewing it much more closely. Learned my lesson on that one.

I'm doing some research on my paternity leave. Pardon me for not getting into details, but I need to vent somewhere.

Manager wanted me to look into a particular data source. Turns out this data source actually has two formats. We want the more compressed format. Turns out nearly all solutions for parsing and displaying this data are cloud/server based and we, for reasons, need to stay on the desktop. Local data only. Closed network most likely.

Maybe my google-fu is weak. Maybe LLMs have ruined everything. Even manually searching package managers for anything promising just turned up the same 5 year old, unmaintained project, and multitudes of abandoned forks, for working with this preferred format on desktop.

So anyways, I set to work. Converting the project to a modern dotnet format, importing the correct nuget packages versus manually linking random DLLs, updating packages often at least 5 major versions. Many packages I had to find replacements for entirely, they'd long since been abandoned and hadn't kept up with changes in the data format.

I think the thing that gets me most, is how many times this project converts data from A to B to Z to X and then finally back to B for the final user display?! Fucking WHY?

You know another thing that gets my goat? All the double precision drawing primitives in dotnet are in the windows only WPF presentation library. Makes no damned sense. Another thing I want to replace at some point with platform agnostic bits.

That setup I talked about wanting last week with adding a 3090 to my rig for added VRAM headroom? Yeah, I splurged and went for it and due to various snafus, such as janky motherboard cables on the initial PSU and bloody Odindamned direct signature required FedEx shipping, I didn't get my rig properly equipped until tonight. Loading the LLM on PCIE 4 4x sure takes its sweet time but on the back end, I just tested it with a 24b model at 8_0 and it was acceptably fast even before tweaking my settings any. Dropping a 3090 ti in (after dealing with the fiddly 3 pci-e power adapter) just... worked, at least for the LLM part, though now I've gotta fix up my games and tell them all to use the 5080, but still. Time to play with some bigger models! When I'm finished moving house, at least...

Botched several connections on v0.1 of the educational robot mainboard (thank you Chinese display board for having your FPC pins start with 1 on the right), but got enough to implement the inter-circuit comm UART, negotiation, and watchdog behavior. Next big push is trying to get over-wifi debugging running; it's theoretically plausible and realistically a nightmare with warnings like "don't use the second core of this processor defined by its multicore capability". Got some fixes in, and trying to figure out if I'm willing to deal with tuning my own NFC sub-board before I order v0.2.

Trying to throw together an agentic layout for editing, proof-reading, summarizing, pulling style notes from, and continuing a story. There's a lot of attempts, here, at varying levels of success, but the state of the ecosystems is just absolutely terrifying, and that's the '''good''' stuff rather than OpenClaw. Not just the obvious but hard problems like 'how do we set up permissions and access rules', though it's disappointing that the answer to that is near-universally 'docker' at best. I get the core idea of content-as-code, but holy shit people. Just because you want your LLM to be able to modify your template files doesn't mean they need to sit in the same directory as your output files!

I finished the novel last Tuesday and took a week off before editing. I'll start today.

Congratulations! But I would recommend waiting longer before looking at it.

Why? The first chapters were written a long time ago, in early spring.

In On Writing, Stephen King recommends leaving your first draft to rest for at least six weeks before picking it up again. This gives you enough distance from it to assess it objectively: it sort of feels like a novel someone else wrote.

So more work on my ongoing quest for AI building in Minecraft.

Exposing primitives via MCP resulted in too many tool calls. Claude could build some neat stuff but it was very slow. Also a bit prone to making mistakes.

I thought I could fix things by adding in a service to use prebuilt schematics. I found an archive of schematic files, generated isometric images from them, had a local gemma3 classify them, and set up an elastic search service so the MCP could search for models by tags.

It worked pretty well for things like "generate a medieval village" but the gemma3 tags weren't great at providing useful info and I'm probably going to have issues if I try to redistribute those schematic files.

Also it's fundamentally limited, I'm going for structures I've never dreamt of.

So I figured that LLMs are good at writing code.

So I'm currently working on using Starlark scripts to generate nbt files. It's basically a flexible component system. Local only, I haven't wired it into the MCP yet.

Claude and Codex are pretty good at looking at an existing schematic and generating a script for a similar building. Results are a bit more mixed when given only simple instructions.

To really get it working I'll need to add in a lot of validations about wether you can path from the entrance to a room / floor. Then for MCP integration I'll need to have a whole system about returning errors and explaining problems.

I'll need to drop some kind of major info dump on the llm that explains how to use it all.

One day a kid will be able to type "generate an epic murder dungeon and teleport my friends there" and my goal will be complete.