Claude Code Tips to Save Tokens: Context, Compaction, Scope

Each feature loads differently. CLAUDE.md loads in full and is paid for on every request. Skills cost only their descriptions until one is used. MCP servers cost their tool names, because tool search defers the full schemas until a tool is called. A subagent runs in a separate window, and a hook costs nothing unless it returns context.

Size is not the only cost. The docs are blunt that performance degrades as the window fills: Claude starts to lose track of earlier instructions. So the habit is the same whether you care about usage or quality. Run /context to see the split as a coloured grid with suggestions, and /usage for the session's token usage and activity. /cost and /stats are now aliases of /usage.

/compact replaces the conversation with a summary when you need continuity. Give it a focus, such as /compact focus on the auth bug fix, and the summary keeps what you chose instead of what the automatic pass guesses. Compaction is itself a large request, because it reads everything it summarises. Auto-compaction runs the same pass as you approach the limit, and /autocompact sets how full the window gets before it does. To condense only part of a session, press Esc twice or run /rewind, pick a message and choose Summarize from here or Summarize up to here.

Know what survives. Project-root CLAUDE.md and auto memory are re-read from disk. Path-scoped rules and nested CLAUDE.md files return only when Claude touches a matching file again. Up to five of the files Claude read or edited are re-read, most recently modified first, and invoked skills come back capped at 5,000 tokens each and 25,000 in total. An instruction you only typed in chat lives on only inside the summary, so anything that must last belongs in CLAUDE.md.

Scoping is the cheapest saving of all. A vague request such as improve this codebase makes Claude scan widely, while add input validation to the login function in auth.ts needs a handful of reads. The same thinking applies to tools: MCP tool schemas are deferred by default, an installed CLI such as gh adds no per-tool listing at all, and servers you are not using can be switched off in /mcp.

Hooks run outside the model and cost nothing unless they return context. Used well, they shrink it: the docs' example filters a test run down to its failures before Claude ever sees the output. Used carelessly, a hook that injects a whole file after every edit becomes a cost on every turn. This repo's figure-review hook caps itself at twelve findings and points to the full script for the rest.