By Dan Donovan, @askdandonovan
Shortening your prompts saves about 8%. The thing costing you far more is your whole conversation being re-sent on every message. Here are the 12 checks that fix it.
This is the checklist I actually run when my usage starts disappearing faster than it should. Every command and every quote below comes from Anthropic's own cost documentation, linked at the bottom so you can check any of it yourself.
First, look at where your usage has actually been going. Press d for the last 24 hours or w for the last 7 days once it opens.
/usageThen, next time you finish a job and start something unrelated, run this before you type anything else. That single habit is worth more than everything else on this page put together.
/clearThere's a trick going around where you strip your prompts back and write to Claude like a caveman. No please, no thank you, no full sentences. The claim was that it roughly halves your usage.
JetBrains actually benchmarked it. 86 coding tasks, real spend, measured properly.
Worth knowing: even that 8.5% is a ceiling. They forced the skill on for every single response during the test. Left to decide for itself, it would save the same or less.
Here's why it can't do much. Caveman mode only shortens the words Claude types between tool calls. Your file reads, your diffs, your tool results and your error messages all pass through untouched, because compressing those would mangle your code. Those are most of what you pay for.
So shortening your prompts is optimising the small end.
Anthropic put it plainly in their own documentation.
That's the whole thing. Your session isn't a chat that grows. It's a package that gets re-sent every single time you press enter.
One fair detail so you have the accurate picture: that repeated history is charged at the cached rate, which is cheaper than fresh input. So it isn't full price on every message. It still adds up, because the package keeps getting bigger and it goes again on every message and every tool call.
Roughly in order of how much they're worth. If you only do the first two, you've got most of the benefit.
This is the big one. When you finish a job and move to something unrelated, clear. Anthropic's line on it: "Stale context wastes tokens on every subsequent message."
If you might want the session back later, name it first, then clear.
/renameThen /resume brings it back when you need it.
This one surprises people. /compact has to read the whole conversation in order to summarise it, so compacting a big session is itself a big request.
Use compact when you genuinely need continuity. And when you do, tell it what to keep instead of letting it guess.
/compact Focus on the auth changes and the failing tests/usage attributes your recent usage to skills, subagents, plugins and individual MCP servers, each as a percentage of the total.
It also flags any behaviour eating 10% or more, things like long context or cache misses, with a tip attached to each one. Press d for 24 hours, w for 7 days.
When a session feels heavy, this shows you what's taking up space right now. It's the fastest way to discover an MCP server you forgot about is sitting on a chunk of it.
/contextRun this, look at the list, disable anything you're not actively using.
/mcpWhere a normal command line tool exists, prefer it. Things like gh, aws and gcloud cost you no context at all, because there's no tool listing to carry.
Sonnet handles most coding work and costs less than Opus. Keep Opus for genuinely hard architectural calls and multi-step reasoning.
/modelThat switches mid-session. /config sets your default.
Extended thinking is on by default, and thinking tokens are billed as output. The default budget can run to tens of thousands of tokens per request.
For simple tasks, lower it.
/effortYou can also turn thinking off entirely in /config.
Your first message after a long break misses the cache and reprocesses your whole context from scratch.
The cache lasts about an hour on a subscription. It drops to five minutes once you're drawing on usage credits, and it's five minutes by default on an API key.
Your CLAUDE.md file loads into context at the start of every session, so anything in it is being carried even when you're working on something completely unrelated.
Anthropic's guidance is to keep it under 200 lines. Move the specialised workflow instructions into skills, which only load when they're actually used.
Running a test suite, reading long log files, fetching documentation. All of it produces a wall of output that lands in your main context and then gets re-sent forever after.
Send it to a subagent instead. The mess stays over there and only the summary comes back.
"Improve this codebase" makes Claude go and read everything. "Add input validation to the login function in auth.ts" doesn't.
Vague requests are expensive requests.
Use plan mode before anything complicated, so you approve the direction before it gets built the wrong way. Press Shift+Tab to cycle into it.
And when you see it heading somewhere wrong, press Escape straight away. /rewind or a double tap of Escape takes you back to an earlier checkpoint.
If you use either, shut them down when the work is done.
This page is Claude Code specifically. These two cover different ground.
Every command and quote above traces back to one of these two.