Skip to main content
Code Guide
M05 Intermediate Methodology

Plan Mode

Let Claude plan before acting

PDF
← All cards

What is Plan Mode

Plan Mode is Claude Code’s “look but don’t touch” mode. In Plan Mode, Claude can read, search, analyze, and propose approaches, but cannot modify files, create files, or execute commands that change system state.

/plan # Enter Plan Mode
Shift+Tab # Toggle Plan ↔ Normal
/execute # Exit and move to execution

Why It Changes Everything

Boris Cherny (Head of Claude Code at Anthropic) starts about 80% of his tasks in Plan Mode, letting Claude plan before writing a single line of code. Result: execution is almost always correct on the first attempt.

The reasoning: Claude explores the codebase without pressure to act. It identifies hidden dependencies, secondary impacts, edge cases. The resulting plan is more solid than what Claude would produce in reactive mode.

What Plan Mode Allows

  • Reading files and exploring the codebase
  • Architecture and dependency analysis
  • Proposing approaches and action plans
  • Writing to a plan file (notes, PLAN.md)

What it prohibits:

  • Editing existing files
  • Creating new files
  • Executing state-changing commands
  • Git commits

When to Use It

SituationPlan Mode?
Exploring an unfamiliar codebaseYes
Feature touching 3+ filesYes
Significant refactoringYes
Typo fix or simple renameNo
Quick edit on a known fileNo

Plan-Validate-Execute Pattern

1. /plan
Claude explores and proposes a structured plan
2. Review the plan
Modify, refine, validate the steps
3. /execute (or Shift+Tab → Normal Mode)
Supervised execution of the approved plan

Validating the plan before execution catches 30-40% of issues that would otherwise be discovered after the fact, according to community feedback.

Auto Plan Mode

For projects where risk level justifies systematic planning, an --append-system-prompt configuration can force Claude to present a plan before each action:

Terminal window
# Practical alias
alias claude-safe='claude --append-system-prompt \
"Before executing ANY tool, present a plan and wait for approval."'

This pattern reports a 76% token reduction with better outcomes, because the plan is validated before spending execution tokens.

Integration with Agents

Agents defined in .claude/agents/ can also enter Plan Mode via /plan within their context. A common pattern: a planner agent (Opus, read-only) that passes the plan to an implementer agent (Haiku, mechanical execution).

Enter your email to read the full card and get the complete PDF bundle.

All content is free and open-source. We just ask for your email.

PDF: