The Unexpected Effectiveness of One-Shot Decompilation with Claude
13 days ago
- #Claude
- #LLM
- #decompilation
- One-shot decompilation with Claude has proven effective, allowing significant progress on Snowboard Kids 2 in a short time.
- The workflow involves a scorer to prioritize functions, Claude for decompilation, tools for support, and a driver to manage the process.
- Claude's performance is superior to other agents like Codex, with Opus 4.5 matching more difficult functions than Sonnet.
- Defensive tooling and clear instructions are crucial to prevent Claude from wasting tokens and getting stuck.
- The majority of functions (79%) are within Claude's capability, shifting the decompilation constraint from human expertise to compute resources.
- Future workflows may focus on refining LLM output rather than writing code from scratch, leveraging matches as a base.