Plan or just do it?
Claude Code can plan before it edits, or it can just make the change. The trap runs both ways. For a big change whose approach isn't settled, the wrong answers start editing straight away. For a one-line fix, they wrap it in plans, searches and questions. Learn to size the task first, and the right mode follows.
D3 Claude Code ConfigurationExam guide 3.4 and 3.5About 15 minutes
Start here: hanging a picture, or redoing the kitchen
Think about two jobs at home.
Claude Code works the same way. Direct execution is hanging the picture: Claude makes the change and runs the tests. Plan mode is the kitchen: Claude explores the code and proposes a plan, and nothing changes until you approve it. Three helpers fit around them:
- The Explore subagent is the person you send to measure every room. They come back with one sheet of numbers, not a pile of notes.
- The interview pattern is asking the family what they want before drawing anything: Claude asks you questions first.
- Tests first is marking where the cupboards must go with tape on the floor. Each attempt is checked against the marks, and Claude sees exactly which ones it missed.
- Yes (a typo, one validation check, a known file): direct execution. Just do it.
- No: many files, more than one valid approach, or an architectural decision: plan mode first.
- The plan needs a lot of searching first: send it to an Explore subagent, so only a summary lands in the session.
- Nobody has said exactly what's wanted: the interview pattern. Claude asks before it builds.
- The right answers are known, but fixes keep breaking other cases: tests first, then share the failures.
Words in the question that tell you which side you're on
Just do it
- one, single, a check, a typo, a rename
- “the file is known”, “the format already exists”
- the ticket already states the rule
- “quickest”, “most efficient”
Plan first
- many files, folders, screens or services
- migrate, restructure, replace, split
- “two libraries”, “engineers disagree”, “not yet chosen”
- “nobody has mapped…”, “how should they start?”
Try it: play out the task
Pick a task, then pick how you start it with Claude. The simulator plays out the steps, marks any rework in pink, and shows how much of the main session's context the work used. The diagram above highlights where your choice sits.
The minutes and context figures are illustrative, chosen to show the shape. They are not measurements. The exam guide's point is that plan mode prevents rework on big or open changes, and that a simple, well-scoped change needs no plan.
Things to try: typo + plan mode, then typo + direct execution. Then library move + direct execution to watch the rework pile up, and library move + plan mode, then + Explore subagent to see the context meter drop. Last, CSV export + plan mode: a good plan built on a guess.
The four traps
| Trap | What the wrong option looks like | Why it's wrong |
|---|---|---|
| A plan for a small change | Plan mode, an Explore subagent or an interview for a typo or one validation check | Overbuilt. The same edit, several times slower, and nothing in the plan was open. |
| Just doing a big, open change | “Convert one folder, then roll the pattern out”, “make the change step by step” | Editing first picks an approach by default. When the hard cases show up, the early work is redone. |
| Right helper, wrong unknown | An interview when the unknown is the code; tests first when the approach is undecided; a plan when the unknown is what the user wants | Each helper fixes one kind of unknown. Match it to what you don't know yet. |
| Discovery in the main session | Running long searches in the planning session, then compacting afterwards | The output already crowded the session. The Explore subagent keeps it out from the start and returns a summary. |
Practice: three exam-style questions
Answer first, then open “Think it through”.
Rule to remember
Exam guide 3.4: “Plan mode: large-scale changes, multiple valid approaches, architectural decisions, multi-file mods.” “Direct execution: simple well-scoped changes (single validation check).” “Plan mode enables safe exploration/design before committing → prevents rework.” “Explore subagent isolates verbose discovery output, returns summaries.” Exam guide 3.5: “Test-driven iteration (write tests first, share failures); the interview pattern (Claude asks questions first)”.