Plan mode

Boris Cherny coded plan mode in 30 minutes on a Sunday evening, after spotting the pattern in GitHub issues. Technically it reduces to a single instruction: "please don't code". Claude reads the repository without modifying it until the plan is accepted, and the mode is entered with Shift+Tab in «Claude Code».

The design detail that keeps recurring is what plan mode does *not* do. Swapping the tool set for read-only versions would have broken the prompt cache, since the cache matches on prefix and any change invalidates everything after it. The «Claude Code» team instead kept all tools in place and added «EnterPlanMode/ExitPlanMode comme outils», with a system message explaining the mode. The model can enter plan mode on its own without a cache rebuild. Tool search follows the same logic with deferred-loading stubs for unused «MCP» tools.

Cat Wu and Boris Cherny report that plan mode doubles or triples the chance of success on complex tasks, by aligning the approach before execution. The Claude Code team's own practice goes further: one Claude writes the plan, a second reviews it with a staff engineer's eye, and when an implementation goes off the rails the move is to return to plan mode rather than persist. Addy Osmani places it at the front of spec writing, as read-only exploration before any code is generated. Louis Claxton's AI-native SDLC playbook makes it a gate: the build stage starts in plan mode and freezes `plan.md` before the first line is written.