Working with Claude Code: 3. Modes

In Claude Code, "Modes" refer to the different ways the AI interacts with your file system and terminal. You cycle through these modes using the Shift + Tab keyboard shortcut while in the Claude Code prompt.

Mastering these modes—and knowing how to negotiate a plan—are crucial for staying in control of your codebase.

Here is a breakdown of the different modes and how to use them, followed by a guide on modifying plans.

1. The Three Interactive Modes

As of the latest versions of Claude Code, pressing Shift + Tab cycles you through three primary modes:

A. Default Mode (Normal)

• What it does: Claude proposes file edits and terminal commands, and asks for your permission before executing anything.
• When to use it: This is your standard, day-to-day mode. Use it when you are working on sensitive parts of your codebase, trying out new libraries, or when you simply want to review every change Claude makes line-by-line.

B. Auto-Accept Mode (Fast Mode)

• What it does: Claude will automatically execute file edits and terminal commands without asking for your permission. (Note: For safety, it usually still respects .gitignore and won't run blatantly destructive commands like rm -rf / without a warning, but it will write and edit code freely).
• When to use it: Use this when you are doing repetitive, low-risk tasks. For example, if you ask Claude to "Find all instances of var and change them to let across the project," you don't want to approve 50 individual file edits. Auto-accept lets Claude rip through the task.
• Warning: Always have a clean Git commit before using Auto-Accept so you can easily revert if Claude goes rogue.

C. Plan Mode (Architecture Mode)

• What it does: In Plan Mode, Claude is strictly forbidden from modifying files or running commands. It can only read your codebase, analyze it, and output text. Its sole purpose is to formulate a strategy.
• When to use it: Use this before starting a new feature, doing a massive refactor, or when debugging a complex architecture issue. It forces Claude to think through the entire problem before writing a single line of code.

(Note: There is also a Headless Mode (claude -p "prompt"), which is non-interactive and used for scripts and CI/CD pipelines, as we discussed earlier).

2. How to Modify a Suggested Plan in Plan Mode

If you are in Plan Mode and Claude suggests a plan that you disagree with, do not accept it and do not toggle back to Default Mode yet.

Plan Mode is a conversation. You modify the plan simply by talking to Claude and providing feedback. Because Claude hasn't written any code yet, it is very easy for it to adjust its strategy.

Here is the step-by-step workflow for modifying a plan:

Step 1: Identify the flaw in the plan

Read through Claude's proposed steps. Maybe it wants to create a new file when you'd rather modify an existing one, or maybe it chose the wrong library.

Step 2: Give direct, specific feedback

Reply to Claude with exactly what needs to change and why.

Example of a bad modification prompt:

"No, that's wrong. Try again." (Too vague, Claude might just guess and repeat the mistake).

Example of a good modification prompt:

"Let's modify this plan. For Step 2, I don't want to create a new auth_helpers.js file. Instead, put those helper functions inside the existing utils/auth.js file. Also, for Step 4, let's use bcrypt instead of argon2 for password hashing, as we already have it installed. Update the plan."

Step 3: Ask Claude to output the revised plan

Claude will read your feedback, rethink the architecture, and output a newly revised plan. If you have multiple rounds of changes, just keep chatting until the plan is perfect.

Step 4: Execute the approved plan

Once you are 100% satisfied with the modified plan, you have two options:

1. Toggle out of Plan Mode: Press Shift + Tab to go back to Default Mode, and say: "The plan looks good. Please execute it step-by-step."
2. Execute step-by-step (Best Practice): Even in Default Mode, tell Claude to pace itself: "Execute Step 1 of the plan only." Review the code it writes, then say "Great, now execute Step 2." This prevents Claude from trying to do too much at once and losing context.

Summary of the Workflow

1. Press Shift + Tab to enter Plan Mode.
2. Ask for a plan: "How should we implement the user profile image upload feature?"
3. Claude outputs a plan.
4. You review it and say: "Modify the plan: let's use AWS S3 instead of local storage."
5. Claude outputs the revised plan.
6. You press Shift + Tab to enter Default Mode.
7. You say: "Proceed with the plan."

ai

Back