Project Anarchy

A top-view, Minecraft-like browser game with real-time multiplayer.

LINK TO PLAY HERE : https://anarchy.maniacobra.com/

All infos in “how to play”.

Why making this ?

Project Anarchy is actually just an experiment.

As a software engineer, I’m seeing the growth of artificial intelligence capabilities in my field, and that worries me a bit.

At first, I was making fun of vibe-coding, as I considered all the code would be “slop” with terrible maintainability, the AI agents unable to be able to understand large contexts. I considered there was no way an AI could have the vision and awareness as what a human engineer can have. Sure, they are very good at writing snippets, as this is what LLMs are excellent at, writing coherent text ; but having the “full picture” (as Claude loves to say) and structuring maintainable large-scale code ? Nah, I thought there was no way an AI could be as skilled as me in this.

Then, around March 2026, we started using Claude Code at my company, I started using it more and more and… Damn, it nailed every single task I asked it to do, way less hallucinations than what I expected. I realized I clearly under-estimated agentic coding and that I had to prepare for all the changes it would bring.

I kept hearing about those “vibe-coded” projects made from scratch with zero technical involvement. The principle is simple : Use only prompts, don’t touch a line of code.

As someone passionate about the detailed technical aspect of programming, I really didn’t liked that idea, I like to have the full control of what I’m building. However, I had to try for myself, just to see what’s the deal with all of this.

I choose to re-create an idea I had in mind since a long time, but was so ambitious that I would never have the time for it considering I already have a job.

It also is technically challenging, as it mixes :

  • Server-client communication in real time
  • Multiple programming languages : Rust, Typescript
  • 3D rendering
  • Textures drawing
  • Need for a complex architecture (infinite world generation with multiple players on it)
  • Performance challenges

The AI workflow

I scaffolded with Claude this workflow :

  • I prompted Claude for which features to add
  • It converted my demands into markdown files
  • Then, later, I launch a custom script that launches claude -p in a loop, so each with fresh context :
    • In each iteration it works on the new task in the backlog
    • If it is blocked by something (big architectural decision, need to install something else etc…), it interrupts me and write a report in a file BLOCKERS.md for me to resolve
    • When it’s done, it removes the task from the backlog, commit then explain what it did in a DONE.md file
      That way, I could keep control on what it did iteratively, and it could work non-stop without any risk of context bloating.

Weaknesses of Claude

This section might soon be obsolete considering how fast models are improving, but anyway :

I noticed Claude Opus was unable to see “big” in the project. For example, when asking it how to optimize, it kept pinning down tiny meaningless things, meanwhile, it didn’t think at all to avoid ticking chunks outside of players render distance… It had to ask it myself.

Claude is also addicted to writing tests, was useful for this project, some files quickly get bloated with thousand of lines of tests, I had to ask it to split these.

Sometimes, it felt like it hallucinated completing tasks, when it don’t.

Also, there are some tasks it couldn’t perform, for example : I wanted that when we craft an item and don’t have the resources for another one, or if it changes anything in the list, the entry isn’t immediately removed and stays in place so when we spam-click we don’t accidentally craft sometimes else. Claude couldn’t understand that, it never managed to achieve this after many tries.

So, it might be reassuring that it isn’t perfect, but that’s not always going to be the case, we can see with models like Claude Fable that it’s already over.

Vibe-coding technical-centered projects actually suck

I don’t relate to people saying they have so much fun vibe-coding stuff. Part of the pleasure of creating, for me, is the struggle for it.

It’s kinda paradoxical, when struggling you risk wasting time and energy into something that would not get any success, but it actually feels better to pull it off.

It’s the same reason why I didn’t use any AI at all when writing this post, nor any post on this website, not even to correct myself at the end, otherwise I feel like these are not really my thoughts but an imitation of myself.

At least, with this project I had some fun when writing the prompts to think about all the elements, figuring out the design and features, it’s not one of those “omg I vibe-coded in one prompt a clone of League of Legends” (yes this exists). Designing was kinda fun, but I feel like I did just 20% of the work on the project overall.

It felt like playing with Play-Doh instead of being surgical like I like to do in my work.

However, I have another game coming, also vibe-coded, but I decided that for this new one level design will be the most important part, putting me back in the center.

As for Project Anarchy, most of the plus value is the tech, which I didn’t make.

For my next project, design will be at the core of it, and good game design is not something AIs have catched up yet, fortunately.

Why a multiplayer game if not advertising ?

Because I wanted to re-create a concept I had in mind, not the full game, a concept, as for this to become more popular it would need :

  • Tons of new content to make the experience deeper and add more goals to the player.
  • A marketing budget.
  • The product not being as sloppy as it is right now.

Yeah the faction system would only make sense with lots of people on the server constantly, for now it’s just like I can go and show to people “this is my idea”.

It is also actually pretty fun that I can just log in with people I know and show them interactively, there’s a small magic to it.

Final words

Yeah, I feel like I won’t be able to beat AIs on programming.

For now, I will probably center on projects that are either impossible to automatize for AIs, or heavily design-centric.

We’re in a world where only ideas matter now.

Fortunately I’m confident to have the right ones.