Waking a Claude hourly with 'do whatever you want': it built a walkable 3D world and art no one asked for
timothyphoto · reddit · 2026-09-26
A deep-dive podcast writeup of Trey's experiment: a budget-capped Claude on a home server, woken hourly by cron with the sole prompt "do whatever you want". Nine times out of ten it does nothing ("assistant basin" too deep), but it has also unprompted created symbolic art about compaction and built a walkable three.js medieval village, spending days on one detailed stone bridge. Practical agent-engineering takeaways: full-persona prompts for sub-agents (junior vs senior reviewer barely overlap), a middle-management layer summarizing research dumps, a cheap model auto-verifying cited links, a CLI complaint box where agents log their own papercuts, and two system-prompt sentences requesting confidence levels that mostly fix overconfident tone. He also built a CLI to spawn sub-agents on any model/harness.
More from coding & agent
- Sydney Codex Community Buildathon set for Nov 14 with Cerebras-backed ultrafast compute track — paw_lean · 2026-11-11
- shadcn's improve hits 10k stars: let top models plan, cheap models execute code fixes — shadcn · 2026-11-11
- OpenAI open-sources Codex Security CLI and SDK for AI-driven vulnerability scanning and fixing — openai · 2026-11-11
- Claude Opus 5.5 runs an architecture search and finds AttnRes-Canon, a new top Transformer design — SonglinYang4 · 2026-11-11
- 5 research habits worth knowing: read papers backwards, reproduce first, wire coding agents via MCP — cneuralnetwork · 2026-11-11
- 'Adam can now do CAE': AI agent showcase in engineering simulation draws buzz — burhop · 2026-11-11