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
Jump in the discussion.
No email address required.
Notes -
So in an effort to update my skills and move in on the AI hype I've been doing a MCP (Model Context Protocol) integration project.
I couldn't think of a useful and functional project to do, so I'm building an api into Minecraft and integrating it with Claude Desktop.
So far it's pretty fun. The limits of Claude have been interesting.
There must of have been a huge amount of work fine tuning models to get them to work with images.
I assumed it would be able to complex shape generation by reading a region into a 3d array and modifying it. Claude seems to have a lot of problems manipulating a big chunk of json data like that.
I asked it to generate a small house in front of me, and this is what it gave me: https://imgur.com/a/CDNKm7K
Also it was behind me. It seems to have a lot of trouble understanding the direction I'm facing. There might be a problem with the data I'm sending.
It seems to handle the more simple tasks I built primitives for well. Creating cuboids of a block type, messaging players, finding players, spawning creatures.
Lately I've been trying to Dockerize it for easy deploy. I'm not all that experienced with Java so it's been an adventure. Or Docker really, there must be easier ways to debug Docker failures than what I've been doing.
I'm looking forward to doing a deploy and terrorizing my friends with AI commands. I see potential for it in twitch streams.
Longer term I'd like to introduce some features for better building generation. I kind of assumed that the currently available tools would be more advanced given how fast everything else seems to move. But after reading up on Model Synthesis and Wave Function Collapse it seems like it'll be a ton of work to get anything working.
I can probably get it to spawn in things from .structure files, but that's much more limited than what I was hoping for.
Do you mean Claude Code ? If you're running a mcp/code setup then claude code will perform noticeably better. In absence of Claude code, I'd try to use Cursor's agent mode.
Yeah, large jsons cause context rot
For any transformation, it's always good to write a strategy document and define a few unit-tasks that the model can orchestrate together. For building houses, asking it to write a blueprint or put foundational blocks on first, may work better. If you can break down your primitives further, that's always good. They don't need to be explicit MCP/tool-calls. A simple prompt stuffed reference will do.
Lastly, creating typed intermediate structures with better semantic flow helps. Json schemas are wonky, I like pydantic (java probably has something similar) for creating large structured schemas that can be validated in real time. Pydantic captures the semantics better than json. It also allows the model to make edits deep into a nested dataclass, without fear of a breaking change else where. This way you build out the entire house in the intermediate structure. And then relegate the intermediate representation -> projection task to a post processing step.
As long as we're talking about data structures, please use dataclasses for storing data and raw functions as transformations. We are in the year of the lord 2025, and OOP should not be used.
Don't know your familiarity with coding, but Java isn't a good place to start. I would recommend python, but most times it's good to just go out there and do things. Getting stuck choosing between tools is never productive. So, maybe ignore me.
Ok so coding is the part of this project I know, I'm just not super familiar with all of the Java tooling. Unfortunately Java is a necessity for Minecraft mods. The MCP interface stuff is in Python.
Getting a jar file with dependencies into a working Docker container was more of an adventure than I expected.
Nope, Claude desktop is a thing with MCP support: https://claude.ai/download
Using an IDE with an open project to do things unrelated to that project seemed peculiar so I didn't go down that path. I tried out using the MCP in VSCode and it worked well, although the agent did try to re-write my MCP code when it couldn't initially connect. The downside to Claude desktop was that it didn't let me try other llms.
I wasn't aware of the VSCode support, so that's neat.
More options
Context Copy link
More options
Context Copy link
More options
Context Copy link