YouTip

Unlike the web version of Claude, Claude Code is not merely a chat tool, but a **development collaboration system with engineering context, permission control, and execution capabilities**. Therefore, when handling tasks, it introduces three core **thinking modes**: * **Ask** * **Plan** * **Edit** **Understanding the boundaries and responsibilities of these three modes is key to using Claude Code efficiently and safely.** You don't need to manually switch modes in the terminal; Claude will automatically switch. In the VS Code extension, you can manually switch modes: !(#) * * * ## Detailed Explanation of the Three Interaction Modes When using Claude Code in the terminal, you **don't need to manually toggle mode switches**. Claude will automatically determine which mode to enter based on your instruction content. However, **as a user, you must clearly express your intent in the Prompt**, otherwise Claude is likely to behave inconsistently with your expectations. ### 1. Ask Mode: Look but Don't Touch (Read-Only Analysis) **Ask mode is Claude Code's default safe mode.** #### Usage Scenarios When you are in the following states, you should prioritize using Ask mode: * Don't understand existing code * Just took over a new project * Want to locate the cause of a bug, but haven't figured out how to fix it yet * Want to confirm whether certain logic is reasonable #### Mode Characteristics * Claude **can read code** * Can perform analysis, explanation, and reasoning * **Will absolutely not modify any files** * Will not execute any Shell commands You can think of Ask mode as: > **A highly experienced senior engineer with hands behind their back.** #### Example Instructions Explain the JWT verification process in src/auth.ts Where is the database connection initialized in this project? Why might this code have issues under high concurrency? > Professional advice: > > **When unsure whether to modify code, always Ask first.** * * * ### 2. Plan Mode: Plan Before Acting (Planning Only, No Execution) **Plan mode is the most architect-like thinking mode in Claude Code.** #### [](#) Usage Scenarios When facing **complex or wide-ranging changes**, you should use Plan mode, such as: * Adding a core module * Refactoring existing interfaces or business processes * Introducing new technologies (caching, middleware, authentication solutions, etc.) * Modifications involving multiple files and layers #### [](#) Mode Characteristics * Claude will not directly modify code * Will first provide a **complete implementation plan** * Usually presented in the form of a step list (TODO List) * Only after you confirm the plan will it enter the Edit phase for execution The core value of Plan mode is: > **Turning code-first into consensus-first.** #### Example Instructions I want to add a Redis caching layer to the existing API, please give me an implementation plan first Plan how to convert this project's styles from CSS to Tailwind CSS If I want to split this module, what would be the reasonable steps? > Practical advice: > > **The more you use Plan mode, the lower your rework probability.** ### 3. Edit Mode: Direct Execution (Can Write Code) **Edit mode is when Claude Code truly gets to work.** #### Usage Scenarios When you clearly know: * What to change * What to change it to * That the change risk is controllable Then you can directly enter Edit mode. #### Mode Characteristics * Claude will locate relevant files * Generate precise Diffs (differential modifications) * May request to execute commands like testing, building, etc. * All write operations require your confirmation In Edit mode, your role is: > **The final code reviewer, not a bystander.** #### Example Instructions Change the login interface response code from 200 to 201 Fix the type error on line 45 of main.py Refactor this function's readability without changing existing behavior > Professional habit: > > **The more specific your description, the safer Claude's modifications.** * * * ## Core Slash Commands (`/`) Slash commands are used to **control Claude Code's session state and behavior**, not to write business code. Type `/` in the terminal to trigger them. !(#) | Command | Purpose | Expert Tip | | --- | --- | --- | | `/help` | View all commands | Essential for first-time users | | `/cost` | Check current session consumption | Prevent unconscious token burning | | `/compact` | Compress context | Use when conversation becomes long and slow | | `/reset` | Reset session | Very important when switching tasks | | `/docs` | Index documents | Let Claude reference specified documents | | `/review` | Code review | Check Git staging area changes | > Recommendation: > > **/cost and /compact are high-frequency commands for long-term users.** * * * ## A Typical Interaction Flow Example In real development, the three modes are often **used consecutively**. ### Example Flow **1. Ask** Why does my registration interface keep returning 400 errors? _Claude analyzes code and logs, pointing out missing email validation._ **2. Plan** I want to add email format validation and standardize error messages, what do you plan to do? _Claude provides steps: modify validation logic → update tests → adjust response structure._ **3. Edit** Execute according to the plan _Claude modifies files and shows Diff for your confirmation._ > **Ask is about understanding the problem, Plan is about avoiding detours, Edit is about cautious execution.** * * * ## Permission Confirmation Mechanism in Interactions Claude Code is very restrained regarding system operations; any risky behavior will request your confirmation. ### Common Permission Confirmations * **Writing files** Will display code Diff, which requires your confirmation before writing. * **Executing commands** For example, `npm test`, `pnpm build`, will explicitly ask for permission. > **Important reminder:** > > Never confirm without looking at the Diff. > > Claude is powerful, but can still produce errors or hallucinations.