Plan Mode
Plan Mode is a working mode in Claude Code in which Claude refrains from writing code directly. Instead, it first formulates a complete implementation plan and raises clarifying questions; only after the plan is confirmed does it proceed to build the solution.
Core Facts
Claude Code's Plan Mode can be toggled via Shift+Tab.
Claude Code's Plan mode does not directly output code; instead, it first performs structured task decomposition and solution planning, outputting an auditable execution plan.
Claude Code has a Plan Mode in which Claude cannot edit files and can only read files to gather information.
Timeline (last 90 days)
Planning before executing can reduce situations where a wrong direction is discovered midway through writing, making it easier to identify requirement deviations before costly generation steps.
Plan Mode can be enabled by pressing Shift+Tab in the CLI, using a dropdown in the desktop app, or by directly telling Claude.
Claude Code's workflow must include a verifiable artifact, which can be an image or a test case.
Plan Mode essentially just adds one sentence to the prompt telling the model it is currently in planning mode and to hold off on writing code.
Clarifying requirements and boundary conditions through a structured interview before entering development results in lower rework costs the more thoroughly the planning is done.
For complex tasks involving schema design, data migration, and backend construction, one should first use the agent's Plan Mode to output a complete plan for human review before execution.
Frequently entering and exiting OPUS's Plan Mode disrupts cache continuity and increases costs.
Plan Mode 是 Claude Code 提供的一种安全工作模式，在该模式下 AI 可读取文件、分析代码、提出方案，但所有文件写入操作会被拦截，必须由用户显式确认后才能执行
When Codex's Plan Mode is enabled, it only discusses plans with the user without actually modifying files, and executes only after confirmation.
All Facts (17)
Claude Code's Plan Mode can be toggled via Shift+Tab.
80%VerifiedClaude Code's Plan mode does not directly output code; instead, it first performs structured task decomposition and solution planning, outputting an auditable execution plan.
80%VerifiedClaude Code has a Plan Mode in which Claude cannot edit files and can only read files to gather information.
65%VerifiedThe role of developers will gradually shift from 'code writers' to 'intent expressers and quality gatekeepers'.
75%UnverifiedCursor的Plan Mode工作流包括研究代码库、澄清问题、生成可编辑计划和执行构建四个步骤
85%Unverified计划模式体现了AI系统设计中人在回路（Human-in-the-Loop，HITL）原则，OpenAI和Anthropic等机构均将HITL视为防止AI系统产生不可逆影响的核心机制
80%UnverifiedWhen Codex's Plan Mode is enabled, it only discusses plans with the user without actually modifying files, and executes only after confirmation.
60%UnverifiedGemini CLI的计划模式通过Shift + Tab快捷键激活，强制智能体在编写代码前先制定详细计划并等待用户批准
60%UnverifiedGemini CLI addresses AI coding control issues through three key features: Hooks, Skills, and Plan Mode.
60%UnverifiedPlanning before executing can reduce situations where a wrong direction is discovered midway through writing, making it easier to identify requirement deviations before costly generation steps.
50%UnverifiedPlan Mode can be enabled by pressing Shift+Tab in the CLI, using a dropdown in the desktop app, or by directly telling Claude.
50%UnverifiedPlan Mode essentially just adds one sentence to the prompt telling the model it is currently in planning mode and to hold off on writing code.
50%UnverifiedClaude Code's workflow must include a verifiable artifact, which can be an image or a test case.
50%UnverifiedClarifying requirements and boundary conditions through a structured interview before entering development results in lower rework costs the more thoroughly the planning is done.
50%UnverifiedFor complex tasks involving schema design, data migration, and backend construction, one should first use the agent's Plan Mode to output a complete plan for human review before execution.
50%UnverifiedFrequently entering and exiting OPUS's Plan Mode disrupts cache continuity and increases costs.
50%UnverifiedPlan Mode 是 Claude Code 提供的一种安全工作模式，在该模式下 AI 可读取文件、分析代码、提出方案，但所有文件写入操作会被拦截，必须由用户显式确认后才能执行
50%