This project is in beta, AI disclaimer

Here it is in AOE II DE! You can set it up yourself with this GitHub - play against the Rebel Alliance or Empire, and even with friends! You will need AOE II DE, Star Wars Galactic Battlegrounds, and Expanding Fronts.

Something you may not know is that Star Wars is already in the AOE II engine. Star Wars Galactic Battlegrounds (SWGB) came out in 2001 and is a reskin of Age of Empires II to have Star Wars mechanics. They have stormtroopers, sith, TIE fighters, and all other types of amazing things.

This will apply to almost every project I will post from now on. I’ll probably adjust this as things change:

1. Everything I ship will make extensive use of AI
2. I won’t really maintain what I ship, but feel free to file issues
3. I welcome people to fork it and to make communities around the stuff I contribute

If you have a problem with that, sorry! I personally don’t really care and you can choose not to use what I make.

In the last month there has been an absolute explosion of game mods and decompilation progress. Off the top of my head:

• Twilight Princess PC Ports
• Breath of the Wild VR mod
• Gen1 Recomp: a 2.5D mod of Pokemon Red Blue & Yellow
• Halo VR, made by someone who claims they don’t know how to code!
• Mario Kart 64 VR

I’ll be referring to modding as a category that includes things like decomp projects, VR mods, game extensions, new game modes, emulation, etc.

It’s great. However, like all evals, the verdict depends on the tasks - so I’ll describe the tasks too.

Before GPT 5.6: Fable orchestrating GPT 5.5#

Right before GPT 5.6 came out, Fable was available on Claude subscription plans. I already had a $100/month Codex subscription. For Claude, as I only had a $20/month subscription, I could only use Fable in the following way:

• Fable set to low
• Fable using this skill to actually use GPT 5.5 as a researcher & coder

I found this paradigm on Twitter. To figure out if this was better than just using GPT 5.5, I ran two evals:

The receipts to show I’m not capping

First, before you jump in on this, know that they haven’t setup their Stripe integration yet. So even if you make a ton of $, this might be a scam LOL. So it’s not entirely worth it.

Update as of Jul 19, 2026: As of now, GPT 5.6 Sol can do this. Will post some results soon!

I started working on adding raytracing (RTX) to The Legend of Zelda: Ocarina of Time. As in my other posts, I’m generally trying to drive the project via autonomous LLM agents.

I was hoping the project could be as simple as running Claude Opus 4.6 in a loop, saying “implement RTX into this engine, make no mistakes.” I got surprisingly far with this technique. However, I started noticing that it hit a wall, and that it seemed to be bottlenecked by its ability to understand the engine’s screenshot output, identify issues, and figure out what to work on next.

This is probably something I’ll be updating over time but has helped me a ton when kicking off long-running agents.

1. Find a hard problem. Something like creating a browser, adding raytracing to an old game, adding VR to an old game, decompiling an old game (are you starting to see a trend?)
2. Get an agent structure that allows you to run an agent on an always-on machine. It should be the same platform as the machine that will run the “problem” - if it’s a game, probably Windows. I’m using Opencode + Kimaki to check in on Discord
3. Ralph loops. I created a skill here. For the initial codebase, have it create a comprehensive PLAN.md, and then setup a Ralph Loop until the PLAN.md is complete. You can have it keep kicking off agents until there are no - [ ] left in the document.
4. (But not necessarily last): make it so that the agent can somehow verify its output. A screenshot harness, the ability to directly set state to a certain position, are super super important as as your topic gets more niche, AI is less likely to output code that works on the first try.

AI Agents are now magic decompilers. Previously, I ran an agent-in-the-loop to try to decompile Super Smash Bros. Melee in Dec 2024 with gpt-4o, but found that the model didn’t tend to learn from its mistakes. Since then:

• Models have gotten smarter
• Tools have gotten better
• Other people have started to create skills & tools to help AI

It’s not just a couple of functions here or there. I’ve merged around 20 functions, and have 80 more in review. Previously, it would take me 1 day a function. My results aren’t even the most impressive though. The writer of the decomp skills set a record for the most matches in a PR. Tons of people are using their own custom agents or just Claude Code in the Discord (channel is #smash-bros-melee) every day with great results. If this is something you’re interested in, come pop on by! We could always use more tokens 🤠.

Original Article

I’ve felt this tension between “scrubs” and people trying to have fun, especially when playing competitive Smash. I do agree that it’s frustrating when people purport to be competitive but are actually scrubs. There are some points I think are missing though:

• Games can change over time, especially when you complain a ton. Within the last two decades in Melee, wobbling was discovered (2006) and then finally banned (2019). The technique was degenerate, and hated in almost all levels of play. If people just sucked it up and dealt with it, I don’t think it would have been banned. If you’re not familiar, wobbling results in you being unable to move and pummeled until you die for up to minutes at a time.
• Being competitive is great; but it’s not like being competitive and “having fun” are more or less honorable than each other. When you engage in an activity with someone else, the goal isn’t always to “completely crush them.” Sometimes, it’s just to have fun, right? In the case where some techniques aren’t “fun”, it seems fine to limit yourself from using them.
  • If anything it’s a bit socially awkward if you can’t sense the mismatch in competitive drive and don’t tailor to what your gaming partner is expecting.

Score: 8/10

Hollow Knight: Silksong came out recently, and I got so much fomo as I had never completed the previous entry. I’ve attempted to clear the game 2 times, and just started my third playthrough. As of finishing this entry, I beat the final boss and ended up getting the 2nd best ending.

Hollow Knight is hard. If you only have a cursory interest in video games, I don’t recommend the game. As someone who is mildly masochistic and isn’t exactly good enough for Elden Ring, Hollow Knight is at the perfect threshold in difficulty for me where I can clear bosses after around 10 tries and a crashout in my journal.