Anthropic's own tips to make Claude Code limits last

Anthropic's documented advice comes down to this: clear between tasks, match the model to the job, keep CLAUDE.md and your tools lean, and check /usage to see what is actually using your limit. This post collects only habits Anthropic itself documents, with the source for each, so you are not relying on forum folklore. Anthropic does not publish how many tokens or messages a limit holds, so none of these come with a promised saving.

The habits are from the Claude Code costs documentation, the Help Center article models, usage and limits in Claude Code and the usage limit best practices, as of November 2026. For the reasons behind them, see why Claude Code uses up your limit so fast.

1. Clear between tasks

Anthropic lists this first. "Stale context wastes tokens on every subsequent message", because the whole conversation is resent with each request. The Help Center's test: "if your next prompt would make perfect sense in a brand-new session, clear before sending it."

Run /clear. The Help Center warns that it cannot be undone; the costs docs suggest running /rename before clearing so you can find the session later and /resume it. Anthropic's rule of thumb: "use /clear when starting a new task, and /compact when continuing a long one." The choice is covered in /compact vs /clear.

2. Be careful with /compact

/compact is not free. It "reads the conversation it summarizes, so compacting a large context is itself a large request". You can steer it with custom instructions, for example /compact Focus on code samples and API usage, or by adding a "# Compact instructions" section to CLAUDE.md. When you want a fresh start instead of continuity, "/clear costs nothing."

3. Match the model to the job

"Opus costs several times more per turn than Sonnet, and Sonnet more than Haiku. Spending Opus on routine work is the fastest way to drain a daily limit." The costs docs add: "Sonnet handles most coding tasks well and costs less than Opus. Reserve Opus for complex architectural decisions or multi-step reasoning." For simple subagent tasks, specify model: haiku.

As of November 2026 the default model on paid plans is Opus 5.5 (model configuration docs), so choosing a lighter one is an active step. Change it with /model (saved as the default for new sessions, or press s in the picker to switch for this session only), or start a session with claude --model <alias or name>. Fable models draw on your limits faster than other models (Fable article). See which Claude model uses the least of your limit and how to switch models in Claude Code.

One caution: each model has its own prompt cache, so right after a switch "the next request re-reads the whole conversation with no cache hits" (error reference). A switch therefore costs least when there is little conversation to re-read, such as right after /clear.

4. Plan before big changes

Anthropic suggests asking for a plan before big changes. It describes "plan with Opus, execute with Sonnet", available as /model opusplan, which uses Opus in plan mode and Sonnet for execution. For complex tasks use plan mode (Shift+Tab), press Escape to stop a wrong direction early, and use /rewind to go back.

5. Point at files instead of pasting them

The Help Center says to give Claude a bare path rather than pasting content, and notes that the @ prefix injects the whole file. Anthropic also advises writing specific prompts.

6. Keep CLAUDE.md lean

"Aim to keep CLAUDE.md under 200 lines by including only essentials." CLAUDE.md is loaded at the start of every session, while skills "load on-demand only when invoked", so moving specialised instructions into skills keeps your base context smaller (costs docs).

7. Lower the effort level

"Higher effort uses more tokens" (Help Center). Use /effort (low, medium, high, xhigh, max, auto) or claude --effort. On models that allow it you can also disable thinking in /config, but thinking cannot be turned off on Opus 5.5, Sonnet 5.5, Haiku 5.5 or Fable models, so on those, lowering the effort is the way to spend less (costs docs).

8. Trim MCP servers and tools

Run /context to see what takes space, disable unused servers with /mcp, and prefer CLI tools such as gh where they do the same job. In the Claude apps, Anthropic says "Tools and connectors are token-intensive" and suggests turning off the ones you do not need (Help Center).

9. Delegate and preprocess sensibly

Delegate verbose operations such as tests, logs and documentation to subagents so the noise stays out of your main context, but remember that "The subagent's own requests still draw on your usage." Agent teams use about 7x more tokens than standard sessions when teammates run in plan mode. Hooks can preprocess output, for example keeping only the last 100 lines of test output before Claude sees it.

10. Use the /usage recommendations

Run /usage. On a paid plan it shows what counts against your limits and gives recommendation rows, each covering a behaviour behind at least 10% of recent usage, for example "Trim ~/code/api/CLAUDE.md". Press d or w for the last 24 hours or 7 days. With Claude Code v2.1.305 or later, press f to have Claude Code apply the recommendations it can (costs docs). It also lists your heaviest /loop and other session-scoped scheduled tasks, which matter because such a task "fires on its interval even while the session is idle, sending your full context each time".

Remember the breakdown only covers this machine, while the limit bars include everything on your account (shared limits).

Habits for the Claude apps too

Because Claude Code and the Claude apps share one allowance, Anthropic's general tips help as well: plan conversations, be specific, batch related requests in one message, and use projects, where cached content counts less when reused. If your plan has a session limit, Anthropic suggests you "plan intensive work sessions around your five-hour usage windows" (Help Center). The window starts with your first message (when the 5-hour limit resets).

What is not on the list

Anthropic does not suggest creating extra accounts to get around limits. Its terms for the Claude Code wrap-up allowance name "an attempt to get around usage limits (for example, through multiple accounts or automation designed to exploit the allowance)" as a reason to withhold it (Help Center), and its documents say advertised limits "assume ordinary, individual usage of Claude Code and the Agent SDK" (Claude Code docs). If you have more than one account for a legitimate reason, such as a personal plan and a work Team seat, the point is just to keep track of each. If you hit a limit anyway, the documented options are in hit the session limit and hit the weekly limit.

Seeing whether it works

Headroom is a free Mac menu bar app that shows session and weekly use for each Claude Code account signed in on your Mac. After you change a habit, you can watch how quickly the session and weekly percentages climb compared with before. The number in the menu bar is the average session use across your tracked accounts. Headroom does not diagnose causes, so /usage remains the tool for that. Headroom never sends a prompt, so checking costs nothing against your limits. It reads usage figures Anthropic does not publish as an official interface, so it can break until updated. It is independent of Anthropic and not endorsed by it.