1. Use a plan when the approach could be wrong
A typo, a rename, or a broken test with an obvious cause does not need a plan. Start, and stop Grok if it wanders. A plan earns its time when a reasonable person could pick two different designs. Where to store a file. Whether to add a column or a new table. Which page owns a new button.
The failure mode is familiar. Grok picks a design in the first minute, edits twenty files, and the design was the wrong one. You then pay for a cleanup. Reading a one-page plan first is the cheaper review.
2. Plan mode writes the approach, and waits
Type /plan before the task, or /plan followed by the task. Grok may also ask to enter plan mode on its own. That ask is a real question. If you say no, it stays in normal mode and can edit.
Plan mode switches off the edit tools, except for the plan file itself. It does not switch off shell commands. The session can still read the project and run a command, and a command can still write a file. In the default mode a command that changes something asks you first, and that ask is your guard: while you are still deciding, do not approve a command that changes the project.
The plan is saved with the session, and /view-plan opens it again in that same session. A new session does not open it for you. If the next session needs the finish line, copy it into a file in the repository and point at that file.
There is also a keyboard cycle for modes. It changes. Use /plan if you do not want to memorize keys, and trust the shortcut list inside Grok for the rest.
3. A useful plan is short and specific
You are not looking for an essay. A plan you can approve has five parts, and Grok’s plan file is built for them:
- Why this change is being made.
- The approach it recommends. One approach, not a tour of every alternative.
- The files it expects to touch.
- Functions or patterns already in the project that it should reuse.
- How you will know it worked.
If the “how we will know” section is missing, send it back. A plan with no check is a wish. Edit the plan in the preview, or send it back and say what is wrong. Approving means “build this,” so read the file list before you approve.
Sending it back can be one paragraph. “This plan adds a new export service. I asked for a button that reuses the orders page. Rewrite it so the only new code is the button and one test. The API stays as it is.” You are correcting the approach while it is still a page. That is cheaper than correcting it after the files have moved.
4. Cut scope into three piles, then build one thin path
Ask for three piles. Must work. Nice to have. Later. Freeze the must-work pile before the build starts. New ideas during the build go into Later, in the plan file, not into the edit that is already running.
Then ask for a thin slice: one path all the way through, even if it is narrow. One button, one happy path, one test. A thin slice shows that the pieces connect. A wide half-finished layer hides the gap until the end.
5. Let it interview you before the plan
You do not have to arrive with a perfect spec. This prompt is enough:
Interview me about this change, one question at a time.
When you can write a plan, write it.
Do not edit project code until I approve the plan.
Answer the questions in plain language. Then read the plan. The habit after this one is to stop repeating those answers every session: write the project rules once.