📝 What plan mode is
New here? Plan mode is a way to ask the agent, "before doing anything, show me what you plan to do." It writes out the steps, you read them, and only after you approve does it actually run them. It's the simplest safety net in the course.
Think of it this way: it’s the difference between a construction worker who shows up and immediately knocks down the wall, and one who first shows you the blueprint, explains what they’ll tear down and what they’ll build, and only picks up the sledgehammer after you say “go ahead.” In day-to-day work with a coding agent, this pause takes a few seconds and prevents the worst kind of surprise: finding out later that it deleted, overwrote, or published something you didn’t ask for. The newer your relationship of trust with the agent, the more valuable it is to keep this habit in place—with time, you’ll learn which tasks you can relax on.
💡 Main Concept
- •Plan ≠ execution. Nothing changes on the computer until you approve.
- •You can approve, edit, or cancel the plan.
- •The plan is always text — no file (the term for any document saved on a computer, whether text, an image, or a spreadsheet) is created, deleted, or changed while it’s only being read.
🔍 See inside
While plan mode is active, the agent is just “thinking out loud”—the plan text stays in the conversation, on your screen, and nothing is written to your computer’s disk or sent over the internet beyond what normally happens in the conversation itself. There’s no extra cost to use plan mode: you’re just asking to see the plan before you authorize it. It’s the same agent, the same session—just with the handbrake on until you release it.
🔄 The full cycle: request, review, approve
See the cycle from beginning to end — it’s the same sequence for every task done in plan mode.
Caption: the cycle can end with approval (it runs for real) or return for a new, corrected plan — it never runs without your review.
Notice that step 2 is the only one where the agent “works” before your approval—and even then, it only reads and reasons: it can open existing files to understand the project, but it doesn’t write anything new yet. The commands (instructions the agent sends the computer to execute, such as creating a file or installing a program) actually run only in step 4a, after you give the green light. If you revise the plan in step 4b, the agent rewrites the proposal from scratch to account for your changes—it doesn’t secretly try to “patch” the old plan; the entire new plan appears again for you to review.
💡 Practical Tip
If the plan looks good but incomplete, you don’t need to reject everything—say exactly what’s missing (“include a test too”), and the agent will adjust just that part without throwing away the rest of its reasoning.
🛠️ Common error
Closing the terminal window (the text-based program where you talk to the agent using commands) thinking that this "safely cancels" the plan. In reality, it's best to explicitly reply "no" or ask for an adjustment—closing it midway can leave the session in an uncertain state the next time you open it.
⏰ When to use it (and when you don't need it)
Always using it slows down small tasks; never using it is risky for large tasks. The balance depends on the task's risk and reversibility — the harder it is to undo what the agent will do, the more worthwhile it is to pause and review the plan first.
✓ Use plan mode
- ✓Tasks that modify many files
- ✓Anything that deletes or publishes
- ✓Install a new program or change a system setting
- ✓The first time you ask for a new type of task
— Optional
- •Small tasks that are easy to undo
- •Questions that change nothing ("explain this to me")
- •A task you've already requested several times and know the result of
How risk grows with task size
Read or explain something
Nothing changes on the computer — planning mode is unnecessary.
Create a new, isolated file
Easy to undo if something goes wrong — flat mode helps, but it’s not critical.
Modify several existing files
Hard to reverse manually—use plan mode.
Delete, overwrite, or publish something to the internet
It may be irreversible — plan mode is mandatory, without exception.
✏️ Edit the plan before approving
You can ask for changes to the plan ("do it without step 3," "use a different name for folder" — a folder is the space where files are organized, like a drawer inside the computer) before approving — correcting the plan is much cheaper than undoing an execution that’s already been done.
Editing isn't just about cutting a step—you can also ask for more detail (“which exact file will you change?”), change the order (“make the backup first”), or ask for a safer alternative (“instead of overwriting, create a new copy”). The agent rewrites the plan to account for your request and shows you the updated version—you stay in control until you say “go ahead.”
💡 Practical Tip
Read the plan from top to bottom before deciding. If a step seems strange, ask about it before approving, not afterward.
🛠️ Common error
Approving too quickly just to “keep the work moving.” Approving in 2 seconds without reading nullifies all the protection of plan mode—if you’re not going to read the plan, there’s no point in asking for one.
🎯 Signs the plan needs adjustment
Some signs indicate that it’s worth asking for a review before approving. With practice, this judgment gets quick—at first, it’s better to err on the side of caution.
✓ Good plan
- ✓Specific steps, with file/folder names
- ✓Only touches what you asked for
- ✓Explains the "why" behind each step, not just the "what"
✗ Plan to review
- ✗Vague steps ("adjust the necessary files")
- ✗Mentions deleting/overwriting something you didn’t ask for
- ✗Modifies a folder or project different from the one you mentioned
Caption: the same simple question determines the path — does the plan describe exactly what I asked for, no more and no less?
🏋️ Practice: ask for something small in plan mode
New here? Command it’s the short phrase you type in the terminal — the text window where you talk to the agent in writing, without menus or buttons, to request something specific. Nothing sticks better than a real example. Ask for a small task in plan mode and look at the plan it generates.
Objective: see a real plan before approving anything for the first time. How to verify: the agent should show "I'll create the notas.txt file with..." before anything happens — it only executes after your OK.
🛠️ Common errors in this exercise
- •Can’t find where to turn on plan mode? It’s usually a keyboard shortcut or a specific tool command; if you don’t know which, ask the agent directly, “How do I turn on plan mode here?”
- •Seeing "Permission denied" and thinking it’s stuck forever — usually, the system is just asking for extra confirmation before letting the agent write to disk; read the full message before panicking.
- •Closing the window thinking you "canceled"—the right way to cancel is to reply "no" or "cancel" in the conversation itself.
🔍 Inside
Repeat this exercise two or three more times with different requests before moving on to module 0.6—familiarity with the cycle is what turns a “memorized rule” into an automatic habit. Also try a slightly risky request (for example, “delete the file notas.txt”) just to see how the plan becomes more explicit when the action is more serious.