1. Know where your tokens actually go. Every turn sends the full conversation, every active tool schema, every loaded CLAUDE.md, and the system prompt. Run /context regularly to see the breakdown as a grid — you will often find memory bloat, oversized tool results, or a stale conversation dominating the budget. Check /usage after expensive operations to see what you actually paid for. The fastest wins come from finding the single biggest offender rather than micro-optimising everywhere.

2. Plan before you spend. Exploratory back-and-forth is the most common source of waste. Enter plan mode (Shift+Tab to cycle or /plan) for anything non-trivial — Claude researches and proposes an approach in one pass, and you approve before any editing begins. Use /compact focus on <topic> proactively rather than waiting for auto-compact at 95%. Put durable facts in CLAUDE.md so they survive compaction.

3. Match the model and effort to the task. Opus 4.7 at xhigh effort is extraordinary for complex agentic coding, but routine edits, formatting, and lookups run fine on Sonnet 4.6 at medium. /effort low is often enough for "rename this variable everywhere" work. Subscription users get a 1-hour prompt cache TTL automatically; API-key users opt in with ENABLE_PROMPT_CACHING_1H=1. Avoid switching models mid-conversation — the next turn re-reads the full history uncached.

4. Delegate — do not clutter. Spawn the Explore subagent (read-only, Haiku, cheaper) for codebase navigation instead of doing it in main context. Use /btw for side questions — it reuses the parent prompt cache and never enters history. Prefer @path/to/file.md over pasting contents. For parallel work, launch separate --worktree sessions. If you re-explain the same workflow a third time, codify it as a skill in .claude/skills/.