token-shieldlisted

# Token Shield A playbook for running Claude Code cheaply. The core insight: you pay for the same context over and over. Every API call resends everything the session has accumulated, so the levers are (1) make the resend cheap via caching, (2) make the context small, (3) make the model match the job, and (4) make the output short. In that order of impact. Rule zero: measure, then act. Every claim below is either sourced to first-party documentation or marked as a habit. Where this playbook cannot verify something, it says so instead of guessing. ## The cost model Verified against code.claude.com/docs/en/prompt-caching and platform.claude.com/docs/en/build-with-claude/prompt-caching, read 2026-08-12. 1. Caching is a prefix match. The match is exact, so a change anywhere in the prefix recomputes everything after it. There is no per-file or per-segment caching. 2. Requests are ordered so the stable content comes first: system prompt (core instructions, tool definitions, output style), then project context (CLAUDE.md, auto memory, unscoped rules), then the conversation. A change to the conversation layer leaves the two layers above it cached. A change to the system prompt invalidates everything. 3. Cache writes bill at 1.25x base input for the 5 minute TTL and 2x for the 1 hour TTL. Cache reads bill at 0.1x. A session that does not disturb its prefix pays roughly a tenth for everything it re-reads. 4. The TTL resets on every cache hit, so the cache stays warm as long as you