The Effort Dial
Which level to use for which job — and how to stop burning your limit on easy work
Five levels, one command: /effort
The Idea
For a long time the way to make Claude think harder was magic words — "think", "think hard", "ultrathink" — and everyone swore by them. That's folklore now. Anthropic replaced the whole thing with an actual control: the effort dial.
thinkthink hardultrathink/effortIt does what it says on the tin — it tells Claude how hard to try. You set it with one command: /effort in Claude Code, then pick a level. Higher levels mean more internal reasoning before it answers — better on hard problems, slower and heavier on your usage. Lower levels mean it flies through the easy stuff without draining your limit.
The skill isn't turning it up. It's matching the level to the job.
The Five Levels
lowThe boring bulk
Renames, tidy-ups, formatting, docs sweeps — anything where the answer is obvious and there's just a lot of it. It flies through and barely touches your limit.
mediumEveryday stuff
Quick questions, small fixes, the back-and-forth of a normal session. Fast enough to feel snappy, enough thinking to not be silly.
highThe proper work
Features, refactors, anything you'd actually review before shipping. This is the level real work deserves.
xhighThe long jobs
Newest models onlyBuilt for work that runs on its own for hours — long agentic tasks where Claude has to keep its reasoning sharp across hundreds of steps, not think ultra-deeply about one. It's stamina, not extra depth.
maxThe ones that hurt
The bug nobody can find, the plan before a big build, the decision you'll live with for months. Claude sits there and properly reasons it through. It's slower, and it's the heaviest on your limit — that's the trade.
And Then There's Ultracode
ultracode isn't a level — it's a mode, and the difference matters. Levels change how hard Claude thinks. Ultracode changes what Claude is allowed to do: it runs at xhigh effort and comes with standing permission to spin up whole multi-agent workflows — fleets of Claudes fanning out over a task, checking each other's work, and reporting back.
It's built for the big autonomous jobs — the overnight refactor, the full codebase audit — and it burns tokens deliberately. Exhaustive is the point. Don't reach for it on everyday work.
The Honest Bit
The newest models already adjust their thinking automatically — they decide, prompt by prompt, how much reasoning a job needs. So on everyday asks you may notice no difference when you move the dial. It's an override, not a magic button: it earns its keep at the extremes, when you know something Claude can't — that this bug has been ruining your week, or that you're one heavy session from your cap.
And no level fixes a vague prompt. Max effort on a bad brief is just a slower wrong answer — the dial multiplies the quality of what you give it, it doesn't replace it.
The Habit
Before you hand Claude a job, ask what it's worth. A batch of renames doesn't deserve the thinking a gnarly bug does — and your limit is one pot paying for both.
/effort low -> before the batch of renames and tidy-ups
/effort high -> before the feature you'll actually review
/effort max -> before the bug nobody can find