16 Feb, 2026
In my previous posts, I described how coding agents could be used to decompile Nintendo 64 games and that one-shot decompilation was very effective. That approach allowed me to make rapid progress on the Snowboard Kids 2 decompilation, with the percentage of matched code quickly growing from around 25% to 58%.
After that, progress slowed dramatically, requiring me to significantly alter my workflow. With those changes, I pushed the decompilation into the ~75% range before stalling out again, this time perhaps for good, though I would love to be proved wrong.
This post describes how my workflow has evolved as the project matured, what helped, and where I’m currently stuck. My hope is that these observations will be useful for other decompilation projects.
Prioritising Similar Functions
Decompilation attempts take time and tokens, so the choice of which unmatched functions to work on matters a great deal. My original approach prioritised functions based on estimated difficulty. A logistic regression model ranked candidates using features like instruction count and control-flow complexity, and Claude would always attempt the ’easiest’ remaining function. That worked remarkably well early on, but it eventually ran out of steam. At some point, everything left was hard. Reordering the queue didn’t magically make those functions easier.
At the same time, Macabeus was exploring function similarity via text embeddings of assembly instructions, which then allowed querying for nearby functions in the high-dimensional latent space. This seemed promising. Claude’s output already hinted that it could recognise similar functions and reuse patterns across them. The intuition here is that decompiled functions provide a useful reference to Claude for how particular blocks of assembly can be mapped to C code.
To test this out, I wrote a tool to compute similar matched functions given an unmatched function and adjusted the agent loop to prioritise functions with similar (matched) counterparts. This approach proved highly effective. There were indeed many similar functions that Claude hadn’t previously been able to identify, and these proved invaluable for helping guide its decompilation attempts.
UMAP 2D projection of function embeddings from 27 December 2025, with some arbitrary modifications to make it fit nicely into a blog post.
Computing Function Similarity
... continue reading