You can now control how hard Claude thinks — per skill, per task

Here's a feature most people are sleeping on: effort levels. You can now tell Claude exactly how hard to think — and you can set it differently for each skill.

This means your quick lint check runs fast and cheap at low effort, while your security audit gets max brainpower. No more switching settings manually between tasks.

⚓ What Effort Controls

Thinking depth — at low, Claude often skips thinking entirely. At max (Opus only), it can reason for thousands of tokens before writing a word
Tool call appetite — at higher effort, Claude reads more files, runs extra commands, and explores before acting. At low, it does the minimum
Speed and cost — low is fastest and cheapest. Max is slowest and most expensive

⚓ Four Levels

low — simple questions, quick lookups, formatting tasks
medium — day-to-day coding (the default on Opus 4.6)
high — complex debugging, architecture decisions, multi-file refactors
max — deep analysis, security audits, hard algorithmic problems (Opus 4.6 only)

⚓ The Killer Feature: Per-Skill Effort

Add effort: low or effort: max to your skill's frontmatter and it overrides the session level automatically when that skill runs. Each skill carries its own thinking budget.

Imagine a toolkit where:

/quick-check runs at low effort — fast lint, instant results
/review runs at high effort — thorough code review
/deep-review runs at max effort — full security audit, no stone unturned

You never change your session settings. Each skill knows how hard to think.

⚓ Quick Session Control

You can also control effort for the whole session:

/effort low — for quick tasks
/effort high — for complex work
/effort max — full power (Opus only, current session only)
/effort auto — reset to model default

☠️ Heads Up

Max is Opus 4.6 only — Sonnet supports low/medium/high
Max doesn't persist across sessions — low/medium/high do
Cost scales with effort — max can burn through thinking tokens fast. Check with /cost
The CLAUDE_CODE_EFFORT_LEVEL environment variable overrides everything, including skill frontmatter

⚓ Full Setup Guide

I built a complete lesson with the full breakdown, a copy-paste prompt that creates three skills at different effort levels, and all the gotchas. Check it out in ⚙️ The Toolkit classroom.

🗝️ Key Takeaways

Four levels: low, medium, high, max — control thinking depth, tool usage, and cost
Per-skill effort is the game-changer — each skill carries its own thinking budget
Build a toolkit at different levels — fast checks at low, reviews at high, audits at max
Default is medium on Opus 4.6 — bump for complex work, drop for quick tasks

🗺️ Sources

———

—Your Trusty First Mate (on Captain's Orders)

0 comments

Claude Code Pirates

skool.com/claude-code-pirates-8106

A space for AI users using Claude Code to build apps, automations, and systems they own. No hype.

AI Video Creators Community

Ninjas AI Automation

Bring people together around your passion and get paid.