02 · Best practice
Use plan mode for anything non-trivial
Plan mode forces Claude to think before it edits. Approve the plan, then implementation gets boring.
Plan mode is Claude Code's "pause before you build" gear. It forces Claude to write down what it plans to do, get your approval, and only then touch a file. For PMs, this is the difference between watching Claude rewrite three files you didn't want it to touch and getting a clean PR.
The instinct to skip plan mode "to save time" is the wrong instinct. You're paying one extra turn to avoid a twenty-turn rework when the agent picks a path you'd have stopped earlier.
How to do it
Toggle plan mode (Shift+Tab in the CLI) or invoke it explicitly: "Plan first. Don't write code."
Read the plan. If it's wrong, push back in prose — don't try to micromanage individual steps. A good correction sounds like "Use the existing lib/mdx.ts instead of writing a new helper," not "step 4 should be different."
Approve. Only then will Claude write.
Plan mode is cheap insurance. Even when bypass-permissions is safe (e.g. inside a Runtime sandbox where everything is reversible), plan mode is about alignment, not safety. Use it.
What good plans contain
- The files that will change.
- The order of operations.
- The trade-offs the model considered and rejected — read these; they're where you'll catch a bad assumption.
- A list of what's out of scope for this change.
When plan mode isn't enough
Plan mode is in-session. It doesn't persist across resets. For multi-session work or multi-person handoffs, write a real PRD (practice 01) — that survives the chat window.
Anti-patterns
- Skipping plan mode "to save time" and then spending 20 minutes reverting.
- Approving a plan you didn't read.
- Treating plan mode output as a contract. It's a starting point; push back hard.
- Using plan mode for trivial one-line changes. Reserve it for anything multi-file.