The Long Tail of LLM-Assisted Decompilation
Summary
This post chronicles the evolving workflow for decompiling Snowboard Kids 2 using large language models. It covers prioritization strategies based on function similarity, compares vector embeddings with handcrafted similarity measures, and details tooling and workflow improvements (Claude skills, F3Dex tooling, permuter removal, worktrees, hooks, Nigel, and Glaude). It also notes the decline in unmatched functions from 157 to 124 and discusses ongoing challenges with large, macro-heavy, and math-intensive code.