Stop Claude Opus 5.5 Token Waste: 4 Configuration Steps

Four Steps to Reduce Claude Opus 5.5 Token Usage

Someone shared a method to prevent Opus 5.5 from wasting tokens during a session. Follow these steps in order, run a routine task, then check the usage metrics.

1. Set /effort to medium

The effort parameter controls the model's reasoning intensity. Keeping it at maximum forces deep reasoning even for simple tasks. Switching to medium prevents unnecessary deep reasoning on straightforward tasks, reducing token consumption.

2. Run /claude-api prompt-audit and Apply the Diff

This command audits the current prompt and returns a diff. Applying the suggested changes makes each request's instructions more concise, significantly lowering long-term overhead.

3. Add Stop-and-Ask Rules to CLAUDE.md

When information is insufficient or requirements are ambiguous, the model tends to guess and continue working, burning tokens. By documenting clear "stop and ask" triggers in the project's CLAUDE.md file, you give the model boundaries.

4. Move Task Lists to TASKS.md , Not Scrollback

Scrollback is the conversation context. Storing task lists there consumes context on every request. Placing them in a separate TASKS.md file means they are read only when needed. This small change yields large long-term savings.

After applying these configurations, run a regular task and check the usage to see the difference.