Chat Modes
Choose Ask, Code, Plan, Visual, or Review mode for each conversation.
Each chat message runs in a mode that controls what the assistant can do. Pick the mode from the dropdown next to the input box before you send.
Mode reference
| Mode | Behavior | Best for |
|---|---|---|
| Ask | Read-only. No file edits or terminal commands. | Learning the codebase, explaining errors |
| Code | Can read, edit, and run commands. Edits show for Keep / Undo. | Implementing features and fixes |
| Plan | Research only. Writes a markdown plan to .shape/plans/. | Designing an approach before coding |
| Visual | UI and styling focused. Follows existing design patterns. Optional in-chat component preview. | Layout, Tailwind, and component polish |
| Review | Systematic investigation of bugs, failures, and security issues. | Tracking down runtime/build errors, auditing auth, input handling, and dependencies |
Ask mode
Safe default for exploration. The assistant reads your project and answers questions but cannot modify files or run the terminal. Use Ask when onboarding to a new repo.
Code mode
The default for implementation work. The assistant can:
- Read and search files
- Edit files, shown as review cards (Keep / Undo)
- Run terminal commands (see AI Settings for auto-run)
- Call
finishwhen the task is done
Switch to Code when you are ready to apply changes. See Reviewing Edits.
Plan mode
Plan mode researches the codebase and saves an implementation plan as markdown under .shape/plans/. Plans include a ## Todos checklist. It does not edit project source files.
When the plan is ready, review the file and click Build this plan (or switch to Code and ask to implement). During build, the assistant maintains a live todo checklist in chat via update_todos.
Visual mode
Visual mode is for UI work. It reuses your project’s component library when present; otherwise it defaults to Radix + Tailwind. Ask it to preview a component in chat before adding it, or tell it to build without stopping. There is no multi-option gallery or Select UI — reply in chat.
Pair Visual mode with Visual Design on the live preview and Design Tools in the editor.
Older Design selections are treated as Visual.
Review mode
Review mode investigates bugs and security-sensitive code: it reads errors, logs, and relevant files before suggesting fixes, and watches for patterns in auth flows, user input, and unsafe APIs.
By default, Adversarial review in AI Settings runs a second critic pass after Review finishes. Turn that off if you want a single pass.
Older Debug and Security selections are automatically migrated to Review — the two were folded into one mode.
Switching modes
Change the mode between messages in the same thread. Earlier context carries over, but tool permissions follow the active mode for each new message.