CLAUDE CODE

Plan before you build

For a change with more than one reasonable design, have Claude write the approach and wait. Approving a plan is cheaper than undoing the wrong edit.

Figure 1: Claude interviews you, writes a plan file, and waits for approval before it edits.

1. Use a plan when the approach could be wrong

A typo or a broken test with an obvious cause does not need a plan. Start, and stop Claude if it wanders. A plan earns its time when a reasonable person could pick two different designs. Where to store a file. Whether to add a column or a new table. Which page owns a new button.

The failure mode is familiar. Claude picks a design in the first minute, edits twenty files, and the design was the wrong one. Reading a one-page plan first is the cheaper review.

2. Let it interview you, then write the plan

You do not need a perfect spec. A useful opening is: “Interview me about this feature, one question at a time, until you can write a plan. Do not edit yet.” The questions are often the requirements you had not written down.

Claude Code has a plan mode for this. It researches and proposes, and source edits stay blocked until you approve the plan. You can enter it with /plan, or with the mode control in the product. The docs also describe starting the session already in plan mode. If the control has moved, the in-product list is the current one.

Recent Claude Code versions start many interactive sessions in auto mode, which lets more actions run under background checks. That is not plan mode. Look at the mode indicator. A mode that skips prompts is still not a reason to skip the diff. Be especially cautious with any setting that turns approval off entirely.

3. Save the plan in the repository

A plan that lives only in the chat dies when the chat does. Ask for a file, often named PLAN.md, with three tiers: what must work, what is nice, and what is a stretch. Freeze the must-work tier before the build. You edit the file. Claude does not get to promote a stretch goal into the core while you are not looking.

If you are building to a written acceptance list, give it that list and ask the plan to map each line to a check. A plan that cannot point at the requirement is a wish.

Layman rule: Approve the approach before the files change. Catching a wrong design in the plan is the cheap version.

4. Cut a thin path, and leave the rest

The plan should name one end-to-end path that works first, even if it is ugly. That path is the subject of a later page. Widening before the path works hides the integration bug under a pile of features.

A one-line fix can skip this whole page. If you are unsure whether the task is one line, it is not one line.

5. What to try next

On the next task that could be designed two ways, ask for the interview, then the plan file, and do not let it edit until you have changed the plan yourself. The build still needs the real file and the real error: give it those, not a paraphrase.