$ man context-wiki/plan-mode

Plan mode is read-only mode. Claude explores your codebase, reads files, researches patterns, and maps out steps before touching anything. No file edits. No code changes. No commands that modify state. Just thinking. In Cursor, you can switch to plan mode to force the agent to analyze before it acts. In Claude Code, Shift+Tab toggles it. The key insight: in plan mode, Claude asks clarifying questions instead of making assumptions. That alone makes it worth using for anything complex.

PATTERN

Use plan mode for anything with multiple valid approaches or significant tradeoffs. Building a new feature that touches many files. Refactoring a data model. Architecting a campaign workflow from scratch. Debugging a problem you do not fully understand yet. Do not use plan mode for simple, well-defined tasks. If you know exactly what needs to change and where, skip straight to agent mode. Renaming a variable, fixing a typo, adding a single field. Those do not need a plan. The rule of thumb: if the task could go wrong in more than one way, plan first. If it is straightforward, execute directly.

PRO TIP

I use plan mode for anything complex. But I do not just say "make a plan." I specifically instruct Claude to tell me which agents it will use, what context each agent needs, what order to run them, and which tasks can run in parallel. That is context engineering in action. You are engineering the plan that engineers the context. When I built the Clay Wiki, the plan mode output told me exactly which files each agent would create, which existing files needed edits, and confirmed that all four agents could run in parallel because they touched different files. Without that plan, I would have launched agents that stepped on each other.

PATTERN

A good plan mode output includes: (1) Files to create, with full paths. (2) Files to modify, with what changes are needed. (3) Dependencies between tasks. Which must run first? Which can run in parallel? (4) Model recommendations for each task. Heavy creative work gets the default model. Copy-paste-and-adapt gets a fast model. (5) Verification steps. How do you confirm everything worked after execution? If your plan mode output is just a numbered list of vague steps, you did not give Claude enough context. Load the relevant files, explain the architecture, and ask again. The quality of the plan depends on the context you provide. Plans are context engineering too.