Tune /effort so depth matches the task, not every turn
Use /effort (low→max) or CLAUDE_CODE_EFFORT_LEVEL so simple edits stay cheap and hard debugging gets deeper reasoning without paying max on every prompt.
The problem
Every turn gets the same depth: trivial renames burn tokens on overthinking, while architecture work under-thinks and needs heavy steering.
The hack
Run `/effort` and pick low / medium / high / xhigh / max / auto for the session. Raise to xhigh or max only for hard debugging or design; keep high for day-to-day agentic coding.
Why it works
Anthropic’s effort blog and power-user guide treat effort as the primary quality/cost dial. The Claude Code team defaults to high and bumps to xhigh/max only when the problem deserves it.
Setup
1. Start a session and run `/effort` to see the current level. 2. For routine edits: `/effort medium` or `low`. 3. For complex agentic work: leave `high` or switch to `xhigh`. 4. For frontier debugging/architecture: `/effort max` for that session only. 5. Optionally export `CLAUDE_CODE_EFFORT_LEVEL=high` in your shell profile.
Discussion
No comments yet. Be the first to say whether this worked for you.