Pair a writing LLM like Claude with a cheap decision model that only returns numbers, such as TypeSafe Jev. Three ways to split the work, why a probability Claude writes down is a self-report and not a measurement, and why a reason written after the fact for a bare number is worse than no reason.
Adding an organization-wide instruction block does add tokens to every prompt, and that objection is correct. A controlled 180-call test on Claude Sonnet 5 shows one brevity line cutting output 23.9 percent and returning 14.2x its own cost, while the block carrying it fails to break even uncached.
Claude Code loads your CLAUDE.md into every subagent except two. Explore and Plan skip it by design, and no setting changes that. So the agents you fan out most widely are the ones that never see your rules. ETH Zurich measured what the file costs on the occasions it does load.
A large context window in Claude Code feels free, and it is the opposite. Every token you load is re-billed on every turn after. The prompt cache that should make that cheap expires in five minutes, turning a tenth-price read into a higher-price write. And accuracy fades as the window fills. Here is the real cost.
Claude Code effort mode looks like a cost dial: turn it down to spend fewer tokens. The official docs say otherwise. Effort is a behavioral signal, not a strict budget, so low effort does not reliably cut spend and can quietly raise it. Here are the five levels, where they stop, and how to set effort with intent.
A surprising Claude Code bill is almost never one big expense. It is four different cost shapes stacked up: a context window that bills every turn, subagents that each cost a fixed chunk, skills that cost almost nothing until used, and caching that can cut the recurring cost or not. Budgeting tokens means knowing the four shapes.