Tip: Don't Overestimate caveman's Savings — Output Tokens Are a Minority of Your Bill

andrew.ooo·medium signal

The caveman skill genuinely cuts conversational output tokens (~30-50% in independent tests), but 80-90% of a Claude Code bill typically goes to thinking and code generation, not prose preamble. Use it for chatty review/analysis loops where output dominates, but pair it with context-reduction tools (code graphs, scoped reads) and thinking-token controls for the parts of the bill that actually matter.