Developer Chris Lewis spent three months manually decompiling Snowboard Kids 2, an N64 game from 1999, and reached 25 percent completion. Then he automated the process with Claude Opus 4.5. Three weeks later, he hit 45 percent. The work processed over 1,000 functions, often matching compiled code in under 10 attempts per function, and topped Hacker News on December 6 with 209 points.
This is not about AI benchmarks. This is proof that Claude can reverse-engineer compiled binaries and recreate source code byte-for-byte—a capability with implications for game preservation, security research, and the billions of lines of unmaintained legacy code rotting in production systems.
What One-Shot Decompilation Actually Means
N64 decompilation is not just reverse engineering. It is a matching process: analyze MIPS assembly from a ROM, infer behavior, write C code that compiles back to byte-identical assembly. Not functionally equivalent—exact binary match. Same registers, same delay slots, same instruction order. The result must look like something an N64 developer in 1999 would plausibly write.
Lewis automated this with a headless Claude Code loop. A scorer prioritizes functions by difficulty using logistic regression. A bash driver feeds functions to Claude, which attempts matches up to 10 times. Successful matches get committed to git. Failures get logged. The system runs unattended for eight hours.
One-shot means Claude analyzes a function once and moves on. No iterative back-and-forth. Just automated single-pass attempts at scale. The workflow targets the simplest functions first—low-hanging fruit—rather than wasting attempts on impossible problems. This strategy unlocked the speed gains.
Why Claude Opus 4.5 Works for This
Claude Opus 4.5 is the first model to break 80 percent on SWE-bench Verified, hitting 80.9 percent compared to Sonnet 4.5’s 77.2 percent. It uses 65 percent fewer tokens than previous models for the same quality. Lewis tested both: Opus matched five of seven functions that Sonnet deemed too difficult.
The model excels at pattern recognition in assembly code. It spots compiler optimizations and infers logical structure from low-level instructions. Token efficiency matters here—decompiling 1,000 functions burns tokens fast. Opus 4.5’s improvements make this economically viable at scale.
Anthropic’s headless mode enables the automation. Command-line scripting, JSON output, no persistent state between runs. The bash loop invokes Claude repeatedly, handles failures with exponential backoff, and logs everything for debugging. Simple tooling, massive results.
The Numbers Tell the Story
Lewis estimates 79 percent of Snowboard Kids 2 functions are matchable with Claude. The automated system processed over 1,000 functions in three weeks. Manual decompilation took three months to reach 25 percent. Automation reached 45 percent in the same timeframe. The gap is not incremental—it is a phase shift.
This is not theoretical. The Snowboard Kids 2 project is public on GitHub. Progress is tracked on decomp.dev. Anyone can verify the results. The work hit the Hacker News front page because developers recognize the practical implications immediately.
Game Preservation Gets Real
Retro game decompilation preserves games before source code is lost forever. Super Mario 64, Zelda Ocarina of Time, and Pokemon have active decompilation projects. Manual work takes years. Claude automation compresses that timeline to months or weeks.
This democratizes reverse engineering. You no longer need deep MIPS assembly expertise to contribute. The AI handles pattern matching. Humans focus on architecture decisions and quality cleanup. Lewis notes future workflows will likely center on “cleaning and documenting LLM output rather than generating byte-perfect code from scratch.”
Game preservation is the visible use case. The invisible one is legacy code. Gartner predicts 40 percent of legacy modernization projects will incorporate AI-assisted reverse engineering by 2026, up from less than 10 percent in 2023. Billions of lines of unmaintained code sit in production. Claude can help recover understanding when original developers are gone and documentation never existed.
AI Understanding, Not Just Generation
Most AI coding stories focus on generation: write new code, build features, automate tasks. This is different. Claude is understanding compiled binaries and recreating source code. Reverse reasoning, not forward generation.
This capability extends beyond games. Security researchers analyze malware. System engineers reverse proprietary protocols. Compliance teams audit legacy systems. The ability to automate reverse engineering at scale changes what is economically feasible.
Lewis is clear about limitations: “while the roles may change, I don’t see human experts becoming unnecessary any time soon.” Opus matches 79 percent of functions. The remaining 21 percent are too complex. AI output often has awkward pointer arithmetic and control flow that needs human refinement. This is augmentation, not replacement.
What Comes Next
The Hacker News discussion shows developer interest is high. Other retro game communities are watching. Expect adoption beyond N64—SNES, PlayStation, Sega systems all have similar decompilation challenges. Claude automation could accelerate all of them.
Broader trends support this. RevEng.AI, Spice86, and other AI reverse engineering tools are gaining traction. Check Point Research demonstrated using ChatGPT for malware analysis. The ecosystem is forming around AI-assisted code archaeology.
Lewis proved one-shot decompilation works at scale. The workflow is open source. The results are measurable. The question is not whether AI can help reverse engineer legacy systems—Lewis answered that. The question is how fast other projects adopt the approach and what gets preserved that would otherwise be lost.
