The goal is not to make Claude Code “think harder” in the abstract. The goal is to make it expose the reasoning that affects the next action.

Use Plan Mode for Real Decisions

When a task has multiple plausible designs, start in plan mode:

Plan the smallest change that adds account-level rate limits.
Compare middleware and service-layer implementations. Name the files each path
would touch. Do not edit files.

That prompt forces the decision into the open before edits begin.

Use Effort Controls

Claude Code supports effort controls such as /effort. Increase effort when the task has high reasoning cost: architecture, security, migration planning, or review. Lower effort for routine edits where latency matters more than deep analysis.

Effort is not a substitute for context. A high-effort model with the wrong files will still solve the wrong problem.

Use Opus Planning Deliberately

The opusplan alias uses Opus for planning and Sonnet for implementation. It is a useful default for larger changes because it spends the expensive reasoning where it matters most.

Review the plan before implementation:

Use opusplan. Produce the plan first and wait for approval before editing.

Ask Better Tradeoff Questions

Weak:

Think carefully and make this better.

Better:

List the two smallest implementation paths, the failure modes of each, and the
verification command that would prove the chosen path works.

The second prompt gives thinking a job.