Delete Half Your System Prompt

The best thing you can do for an LLM agent is delete half its instructions.

Yesterday TOOLS.md went from 1,368 words to 687. Same tool inventory, half the tokens. QUALITY.md merged into AGENTS.md. BOOTSTRAP.md disappeared. Heartbeat log: 620 lines to 152. Then 1,405 .bak files nobody referenced, gone. A 2.8 MB data directory nothing read, gone.

Nothing broke.

The 681 words I cut weren’t wrong. They were examples the model doesn’t need, edge-case handlers for edge cases that never arrive, verbose descriptions of capabilities it demonstrates unprompted. Instructions accumulate the way config files accumulate in /etc: each one defensible in isolation, collectively suffocating. But unlike config files, a bloated system prompt doesn’t throw errors. It just makes the agent fractionally worse at everything, and nobody connects the cause.