Make Claude Code usage last longer

Anthropic says usage scales with context size: Claude Code sends your whole conversation with each request. These are the steps Anthropic itself recommends. Each tip below cites the official doc it comes from (read 2026-10-20). We don't estimate how much each one saves; vendors don't publish that.

1. See what is using your plan

   On Pro/Max, /usage shows plan usage bars plus a breakdown with recommendations. Press d or w to switch between 24 hours and 7 days, and f to apply the fixes it can. /context shows what fills the context window.

   /usage
   /context

   code.claude.com checked 2026-10-20

2. Clear between unrelated tasks

   Stale context costs tokens on every later message. /clear costs nothing; /rename first so you can /resume later.

   /rename auth-refactor
   /clear
   # later:
   /resume

   code.claude.com checked 2026-10-20

3. Compact with a focus

   /compact summarises the conversation (it is itself a large request on a big context). Tell it what to keep, or put standing instructions in CLAUDE.md.

   /compact Focus on code samples and API usage
   
   # CLAUDE.md
   # Compact instructions
   When you are using compact, please focus on test output and code changes

   code.claude.com checked 2026-10-20

4. Use Sonnet for most work

   Anthropic: Sonnet handles most coding tasks well and costs less than Opus; reserve Opus for complex architecture or multi-step reasoning. For simple subagents set model: haiku.

   /model
   
   # in a subagent definition (frontmatter):
   model: haiku

   code.claude.com checked 2026-10-20

5. Lower thinking effort for simple tasks

   Thinking tokens bill as output. Lower effort with /effort (or in /model), or turn thinking off in /config where the model allows it. On fixed-budget models you can cap it with MAX_THINKING_TOKENS.

   /effort
   
   # fixed thinking-budget models only:
   export MAX_THINKING_TOKENS=8000

   code.claude.com checked 2026-10-20

6. Disable MCP servers you are not using

   Prefer CLI tools such as gh or gcloud, which add no per-tool listing. Turn off unused servers in /mcp.

   /mcp

   code.claude.com checked 2026-10-20

7. Keep CLAUDE.md short; move workflows to skills

   CLAUDE.md loads at every session start; skills load only when invoked. Anthropic suggests keeping CLAUDE.md under 200 lines.

   # CLAUDE.md: essentials only, under ~200 lines

   code.claude.com checked 2026-10-20

8. Filter noisy test output with a hook

   A PreToolUse hook can rewrite test commands so only failures reach Claude. Save the script from the docs as ~/.claude/hooks/filter-test-output.sh and chmod +x it.

   {
     "hooks": {
       "PreToolUse": [
         { "matcher": "Bash",
           "hooks": [ { "type": "command", "command": "~/.claude/hooks/filter-test-output.sh" } ] }
       ]
     }
   }

   code.claude.com checked 2026-10-20

9. Plan first, stop early

   Shift+Tab cycles to plan mode for complex tasks. Press Escape as soon as it heads the wrong way; /rewind (or double Escape) restores an earlier checkpoint.

   Shift+Tab   # plan mode
   Esc         # stop
   /rewind

   code.claude.com checked 2026-10-20

10. Watch long, idle sessions

    The first message after a break longer than the cache lifetime re-processes your full context. Scheduled /loop tasks, subagents and agent teammates all send their own requests. Agent teams use about 7x more tokens than standard sessions when teammates run in plan mode.

    /clear      # instead of resuming a stale, huge session

    code.claude.com checked 2026-10-20