December 6, 2025
Roomba for code
The Unexpected Effectiveness of One-Shot Decompilation with Claude
AI “vacuum” chews through old game code while you sleep—fans cheer, skeptics fume
TLDR: A dev let Claude run unattended and rapidly decompile parts of a 90s game, claiming weeks of progress in days. Commenters are split between calling it a game‑changer and warning it only shines with guardrails—plus a spicy debate over whether easy decompiling means an open‑source free‑for‑all.
An overnight AI marathon just blasted through a 90s Nintendo game, and the comment section is on fire. The author set Claude to run in a simple “one-shot” loop—basically a set it and forget it mode that tries a function once, then moves on—and says they did in three weeks what took three months before on Snowboard Kids 2. Cue the cheering section: one fan yelled that if you’re not using AI for reverse engineering, you’re “missing out,” while another called it a rare practical win for AI.
But the drama? Oh, there’s drama. One commenter dropped the spicy question: if decompiling (turning scrambled machine code back into readable code) gets this easy, does everything become open source by default? That sparked a mini moral panic in the thread, while cautious voices chimed in that this only works because the author built smart guardrails: pick easy wins first, bail after 10 tries, verify results, commit. Another warned that AI still can’t magically explain complex code to non-experts—it’s a power tool, not a tutor.
There’s even a model war subplot: the author found Claude’s Opus version beat Sonnet on tougher bits. Jokes flew about a “Roomba for code” (the script’s literally nicknamed “vacuum”), wake-and-commit mornings, and “AI speedrunning” old games. Internet, never change.
Key Points
- •A one-shot, headless Claude workflow significantly increased decompilation throughput on Snowboard Kids 2.
- •Automation is driven by a “vacuum” script and four components: scorer, Claude, tools, and driver.
- •Initial function scoring used a hand-crafted formula; later improved via logistic regression, removing stack size as a feature.
- •Opus 4.5 outperformed Sonnet 4.5 in a brief test, matching 5 of 7 previously too-difficult functions.
- •Prompt rules enforce abandoning after >10 failed attempts and committing matched code to preserve progress.