(Written with AI help, then checked by me against the official Claude Code docs, October 2026. Claude Code changes often, so version numbers are noted.)

Every time you send a message, Claude Code sends the whole conversation again : system prompt, CLAUDE.md, every earlier message and tool result. Prompt caching is what keeps that from being reprocessed from scratch each turn. When the cache misses, the whole history is processed again.

So most of the habits below come down to one idea: don't make the cache miss when you don't have to.

One caveat before the list. Anthropic doesn't publish how each token type counts against subscription limits, so I can't tell you any of this will save you a specific amount. Check /usage before and after a change and judge for yourself.

1. Know your cache lifetime (and verify it)

The cache expires after a period of inactivity.…