if you're hitting your claude limit too fast with openclaw, it's probably not your prompts - it's your config was burning through $200 claude max in 4 days. these changes made it last: 1. default to sonnet. only switch to opus when it actually matters 2. sub-agents for everything — their context is separate and dies after the task. main session stays lean 3. trimmed MEMORY(dot)md and AGENTS(dot)md hard. every token in those files loads on EVERY message. archived stale stuff, kept core files under 2-3k tokens 4. skills for repetitive workflows — fewer back-and-forth = fewer tokens 5. requireMention: true on group chats — without it every message hits the API just to decide "nah." mention gating skips the call entirely 3 and 5 were the biggest wins. people underestimate how much workspace files cost when they're loaded on every single input.