Plan Mode & Extended Thinking

Plan Mode

Plan mode separates thinking from doing. Instead of immediately executing, Claude produces a detailed plan for your review. This is essential for any task with more than 3 steps or architectural implications.

Activate it explicitly in your prompt:

Plan only — do not make any changes yet.
Plan how you'd implement a real-time notification system using WebSockets.

Claude will produce a phased implementation plan, list risks, and ask clarifying questions before touching any code.

When to Use Plan Mode

• Features that touch more than 5 files
• Database schema changes
• Authentication flow changes
• Public API design
• Any work that's expensive to undo

The rule of thumb: if you'd want a code review before merging, use plan mode before implementing.

Extended Thinking

Extended thinking gives Claude additional reasoning budget — it "thinks" longer before responding, useful for complex problems:

• Architecture decisions between tradeoffs
• Debugging subtle race conditions
• Analyzing security implications
• Evaluating competing approaches

Enable it with a prefix:

Think carefully about this before answering: what's the safest way to migrate
our auth system from sessions to JWTs without logging out existing users?

Reading Plans Before Approving

When Claude produces a plan, read it critically:

1. Scope check — is it doing more than you asked?
2. Order check — are steps in the right sequence?
3. Risk check — what could go wrong at each step?
4. Rollback check — can we undo each step if needed?

Push back on plans that seem too broad or risky. It's much cheaper to refine a plan than to undo an incorrect implementation.

Iterating on Plans

Plans aren't final. Treat them as drafts:

Your plan looks good but step 3 is risky. Can you add a feature flag so we can
roll back the new auth system without a deployment?

Refine until you're confident, then ask Claude to execute the approved plan.