nobody audits their Сlaude setup here's what that costs: > your bill creeps > your output drifts > you blame the model here's what actually moves the needle: > 125+ keys exist in Claude Code's settings.json. official docs cover ~40 > fixing one cache_control breakpoint cut a $340/month bill to $87 > unused MCP servers load their full schema on every session > that's 25-40K tokens before you type anything > 14 settings live 3 clicks deep in the Claude UI > 4 aren't documented anywhere the fixes aren't complex: > set Extended Thinking to Light by default - Full only for hard problems > use the `enabled: false` flag on MCP servers instead of deleting them > set `cleanupPeriodDays: 180` - the default 30 buries 5 months of signal > put cache_control after the system prompt, not after the user message > one Custom Style per workflow replaces 80% of saved prompts > scope memory per Project or it drifts and starts optimizing for the wrong "you" the pattern is the same across all 18: the default ships for the average user you are not the average user 20 minutes with a checklist explains why Claude "got dumber" from people who know exactly why it got sharper