Plan Mode is a read-only operating mode in 📝Claude Code where the agent analyzes the codebase, proposes an implementation approach, and surfaces it for explicit user approval before any file writes or destructive actions.
Plan Mode is entered programmatically via the EnterPlanMode tool, or interactively from a user prompt. While active, Claude can still read files, run informational shell commands, and explore the codebase — but Edit, Write, and any mutating Bash operation are blocked. The session resolves when Claude calls ExitPlanMode with a plan body, which the user approves (unlocking writes) or rejects (sending Claude back to planning with feedback).
The pattern was designed for high-stakes changes where a single misstep is costly: architecture refactors, schema migrations, multi-file API edits, security-sensitive code. Forcing the agent to articulate a complete approach before touching anything moves the decision from reactive small edits to reviewable architectural proposal — and produces an artifact useful for archive, human review, or adversarial second-opinion.
A common practitioner workflow pairs Plan Mode with 📝Codex via the 📝Codex Integration for Claude Code skill: Claude drafts in Plan Mode, the operator passes the plan to Codex for an independent review, Codex flags missing constraints or risky assumptions, and only then does the human approve ExitPlanMode. The two-prompt pattern (code review vs. premise audit) distinguishes "does this code work" from "is this the right code to write" — both should run before approval on consequential changes.
Plan Mode is where I slot Codex for a premise audit before any non-trivial build — Claude drafts, Codex challenges the assumptions, and only then does ExitPlanMode get the green light. The friction is the point.