You are three hours into something good. The architecture is finally right, he understands the codebase, and then the message lands. Usage limit reached. He has clocked out, and everything he knew about your project clocked out with him.
Here is the direct answer. Claude Code does not run out of usage because it is working hard. It runs out because of two habits, and Anthropic names both of them: long sessions that were never cleared, and Opus left as the default model. Every message you send re-sends the entire conversation, so a session you never cleared quietly gets more expensive with every single turn. Fix the context and fix the model routing, and the limit stops arriving mid-task. The fourteen commands below do exactly that.
The Full List, at a Glance
| # | Command | What it does |
|---|---|---|
| 01 | /model opusplan | Opus plans, Sonnet builds |
| 02 | /compact | Wipe the garbage, keep the decisions |
| 03 | ultrathink · /effort | Make him actually think first |
| 04 | /clear | New task, clean desk |
| 05 | /context | See what is eating your tokens |
| 06 | /usage · /cost | How close is he to clocking out |
| 07 | /model sonnet | Stop paying Opus prices by default |
| 08 | .claude/agents | Hire interns for the grunt work |
| 09 | Shift+Tab (plan mode) | Look before he touches your code |
| 10 | /rewind | Undo a bad direction instead of arguing |
| 11 | /resume · claude -c | Stop re-explaining your project |
| 12 | CLAUDE.md · /memory | The employee handbook |
| 13 | /mcp | Fire the tools he never uses |
| 14 | auto-compact | He cleans his own desk, on by default |
Part One: The Three That Matter Most
01. /model opusplan
Stop scheduling your most expensive employee for every small thing. opusplan is a model alias that, in Anthropic's exact words, "uses opus during plan mode, then switches to sonnet for execution." Opus, the one who is genuinely brilliant at deciding what to build, runs while Claude is planning. Sonnet, the one who is fast and cheap and types beautifully, runs while it is actually building. You get Opus-quality decisions at Sonnet-level burn, and you do not have to switch models by hand.
/model opusplan
# check what is active at any time with:
/model
02. /compact
Right now, every time you ping him, he re-reads the entire chat from day one. /compact summarises the conversation into a tight brief and keeps going, so you keep the decisions and throw away the noise. The part most people miss is that you can tell it what to protect.
/compact /compact focus on the payment refactor, keep file names and decisions
You can also make it permanent. Add a ## Compact instructions section to your CLAUDE.md telling Claude what to preserve every time it compacts, and you never have to type the focus again. Run /context first, so you know what you are about to throw away.
03. ultrathink and /effort
Before anything goes to production, make him actually think first. There are two separate levers here, and knowing the difference is what makes this tip work.
The keyword. Include ultrathink anywhere in your prompt and Claude Code recognises it and asks for deeper reasoning on that turn. Anthropic's documentation is precise about what happens: it "requests deeper reasoning on that turn without changing your session effort setting," and adds an in-context instruction. Worth knowing: only ultrathink is a recognised keyword. "Think", "think hard", and "think more" are passed through as ordinary prompt text and do nothing at all.
The dial. If you want the effort level itself turned up, that is /effort, which takes low, medium, high, xhigh, max, and ultracode. This is the one that persists across the session rather than one turn. Use the keyword for a single hard prompt, and the dial when the whole session is hard.
# one hard turn > ship the checkout flow to production ultrathink # the whole session /effort high
The economics. Thinking tokens are billed as output tokens, so higher effort costs more per prompt. The saving is not in the prompt, it is in the loop. One properly reasoned answer is cheaper than three confident wrong ones plus the cleanup. Spend it on the hard stuff, not on renaming files.
Part Two: The Other Eleven
04. /clear
New task, clean desk. /clear empties the conversation context completely and starts fresh, and the old session stays saved so you can walk back into it later. Carrying yesterday's debugging session into today's new feature is pure token waste, and it is the single most common reason a session gets expensive. Anthropic's tip is worth stealing: run /rename before you clear, so the old session is easy to find again.
05. /context
The desk audit. /context shows a visual grid of exactly what is filling your context window right now, broken down across the system prompt, tools, MCP servers, files, and the conversation itself, with suggestions on what to cut. You cannot fix a leak you cannot see, so run this before you blame the model.
06. /usage and /cost
Check how close he is to clocking out before he does it mid-task. /usage shows your live plan limits, an attribution breakdown of what has been eating them (skills, subagents, plugins, individual MCP servers), and the session's token and cost figures. Press d or w to flip between the last 24 hours and the last 7 days. /cost is an alias for the same screen, so use whichever you remember.
07. /model sonnet
The always-on version of trick one. Make Sonnet your default and pull Opus in only when a problem is genuinely hard. Anthropic's own guidance says it plainly: Sonnet handles most coding tasks well and costs less, so reserve Opus for complex architectural decisions and multi-step reasoning. Your smartest employee should not be on the schedule for typo fixes.
08. Subagents in .claude/agents
Hire interns. A subagent runs in its own separate context window, which means its long and noisy output never lands in your main session. Pin the mechanical jobs, searching, summarising, boilerplate, to a cheaper model by setting model: haiku in the subagent's frontmatter. The grunt work gets done off your desk, on a smaller salary, without filling your context.
--- name: grunt description: Search, summarise, and boilerplate. The cheap stuff. model: haiku ---
09. Plan mode (Shift+Tab)
Press Shift+Tab to cycle into plan mode and Claude researches your codebase and proposes an approach without editing a single file. You approve before any tokens go into writing code. There is no half-built wrong solution to unwind, because the thinking happens before the typing. This is the cheapest command on the list, and the one people skip most.
10. /rewind
He went down a bad path. Do not spend twenty messages arguing him out of it, because arguing with an AI costs tokens and rewinding is close to free. /rewind, or a double-tap of Esc, rolls both your code and the conversation back to a checkpoint before the mistake. Then give a different instruction and go again.
11. /resume and claude -c
Re-explaining your project every morning is the most expensive small talk there is. claude -c continues your most recent session, and /resume gives you a picker of earlier ones with their full context intact. Pair this with /clear: clearing is not deleting, so you can always walk back.
12. CLAUDE.md and /memory
The employee handbook. Anything you catch yourself repeating, your stack, your conventions, "never touch the migrations folder", goes into CLAUDE.md once and gets read at the start of every session automatically. Edit it any time with /memory. Instructions written once stop being paid for twice.
Keep it short, and here is the number. CLAUDE.md rides along in every session's context, so every line is a line you pay for on every task, even unrelated ones. Anthropic's guidance is to aim to keep it under 200 lines, with only the essentials. Detailed procedures belong in a skill instead, because a skill's body loads only when it is actually used.
13. /mcp
Fire the tools he never uses. Run /mcp to see every connected MCP server and disable the idle ones. Worth being precise here: modern Claude Code already defers MCP tool definitions, so only the tool names enter your context until Claude actually uses one. That is a big improvement, but names are not free, and a dozen servers you never call still cost you something on every session. An unused integration is an employee you are paying to sit in the meeting.
14. Auto-compact
This one is already working for you. When the conversation approaches the context limit, Claude Code summarises it automatically and keeps going. The desk gets cleaned without you asking. The setting is autoCompactEnabled and it defaults to true, so unless you have deliberately turned it off, it is your safety net under everything above.
The Honest Part
Three things this list will not do for you.
It will not make an unlimited plan. The session and weekly windows on a subscription are shared across all models, so switching with /model does not restore access once you have hit the plan limit. It only keeps you working after a model-specific "you have hit your Opus limit" message. What this list actually buys you is arriving at that ceiling far less often, which is a different and more useful thing.
Higher effort is not free. ultrathink and a high /effort spend more thinking tokens on every prompt they touch. The whole bet is that one reasoned answer beats three cheap wrong ones plus the cleanup, and on a production change that bet is right. On a rename, it is not.
Subagents and agent teams are not a free lunch either. They save your main context by spending someone else's. Anthropic's own figure for agent teams is roughly 7x the tokens of a standard session when the teammates run in plan mode, because every one of them carries its own context window. Delegate the noisy work. Do not spawn a committee to rename a variable.
He was never lazy. He was re-reading a three-hour conversation before every reply, in the wrong model, with tools he never used loaded in.
If You Only Do Five Things, Do These
- 01Set
/model opusplanright now. It is one command and it fixes your single biggest source of burn. - 02Run
/clearbetween unrelated tasks. Not at the end of the day. At the end of the task. - 03Run
/contextonce today and actually look at it. You will find something loaded that you forgot about. - 04Use plan mode before anything that touches production. Approving a plan is cheaper than unwinding a build.
- 05Write the
CLAUDE.mdyou keep meaning to write. Under 200 lines. It pays for itself in a week.
Run the list top to bottom once and the naps stop: the right model on the right job, a clean desk between tasks, and actual thinking before anything ships. If you want him to remember what he learned after all this, that is what the memory plugins are for. If you want to stop paying him to read your font choices, stop uploading PDFs. And if you want him to work the way you work, write him a few skills.
Save this one. You will want the commands again the next time he tries to clock out.
Questions People Ask
Why does Claude Code keep hitting its usage limit?
Usually two habits: long sessions that were never cleared, and Opus left as the default model. Every message re-sends the whole conversation, so an uncleared session gets more expensive with each turn. /clear between tasks and /model opusplan fix most of it.
What does /model opusplan do?
It is a model alias that uses Opus during plan mode, then switches to Sonnet for execution. You get Opus-quality decisions while it plans and Sonnet-level cost while it writes the code, with no manual switching.
Does typing "ultrathink" make Claude Code think harder?
Yes, on that turn. Anthropic's docs say ultrathink requests deeper reasoning "without changing your session effort setting", and that the effort level sent to the API is unchanged. To move the dial itself, use /effort (low, medium, high, xhigh, max, ultracode). "Think hard" and "think more" are not recognised keywords.
What is the difference between /clear and /compact?
/clear empties the context completely and starts fresh, with the old session still reachable via /resume. /compact keeps the thread alive but summarises the history into a shorter brief. Clear when switching tasks, compact when continuing one.
How much does Claude Code cost per developer?
Anthropic's documentation reports roughly $13 per developer per active day and $150 to $250 per month across enterprise deployments, with 90% of users staying under $30 on any active day. It varies heavily with model choice and session hygiene.
Is auto-compact on by default?
Yes. autoCompactEnabled defaults to true, so Claude Code summarises the conversation automatically as it nears the context limit. It shows up in /config as Auto-compact, and can be disabled there or with the DISABLE_AUTO_COMPACT environment variable.
Want an AI setup that pays for itself instead of surprising you?
I help founders and brands build generative AI systems that ship: agentic workflows, coding-agent setups, and content engines that are production-grade down to the plumbing. If you want a second pair of eyes on yours, book a session.
Book a Growth Chat
Every command, flag, setting key, and cost figure in this post was checked against Anthropic's Claude Code documentation, current as of July 2026: the cost management page (code.claude.com/docs/en/costs) for the per-developer spend figures and the reduce-token-usage guidance, model configuration (code.claude.com/docs/en/model-config) for the opusplan alias, the ultrathink keyword, and effort levels, the slash commands reference (code.claude.com/docs/en/commands), the settings reference for autoCompactEnabled, and the sub-agents page for model selection. Where the popular version of a tip differs from the documentation, notably around what ultrathink actually changes and how MCP tool definitions are loaded, I have gone with the documentation. Nothing in this post is sponsored.