[Day 11 of 30] Ask First, Plan Second, Then Let the Agent Build
Why the order of modes in your AI coding tool matters
The situation
You open the AI coding assistant, describe a feature in one sentence, and hit go. Minutes later you have edits in a dozen files, some of them wrong or out of sync with how your project works. The real cost is not the first run, it is the rework. The fix is not “better prompts.” It is using the right mode at the right time: questions first, then a reviewable plan, then implementation only after you agree.
The 1% upgrade
Treat your AI coding tool as three gears, not one.
Ask: read-only. Use it to explore the codebase and clarify what you want.
Plan: the assistant researches and proposes a concrete plan (files, steps, trade-offs) but does not edit anything until you approve.
Agent: full execution. Use it only when the plan already reflects what you want. That order (Ask, then Plan, then Agent) cuts rework and keeps you in control. Influencers and docs (e.g. Cursor, Claude Code, DEV community) all point to the same idea: separate planning from execution.
The workflow in practice
Cursor: Open the Agent panel (Cmd+I for Mac or Ctrl+I for Windows). Switch modes by clicking the mode picker dropdown below the chat input (it shows Agent, Ask, or Plan). Use the Shift+Tab hotkey to cycle modes when focus is in the chat input.
Ask is read-only (questions, codebase exploration).
Plan is for complex or multi-file work: the agent builds a plan, you review and edit it, then you run it.
Agent is for doing the work after the plan is approved. For big or unclear tasks, the recommended flow is Plan, then Ask when you need to dig, then Agent to implement.
Claude Code: The mode switcher is in a button below the chat area. Three modes:
Ask Permissions (default), where the agent asks permission before each action;
Auto Accept Edits, which automatically accepts all file edits;
Plan Mode, where the agent can only read files and build a plan, and cannot make changes.
There is also Bypass Permissions, off by default. You can enable it in settings if you would like
For complex or multi-file tasks, the recommended flow is to start in Plan Mode so the agent explores the codebase and proposes a plan, you review it, then switch to Ask Permissions or Auto Accept Edits to run the implementation.
Principle that works everywhere: Before any implementation, write or say what you want in a few sentences (a tiny spec). Ask the AI to break it into steps and list what would change. Review and adjust that plan. Then hand off execution. People who do this often see much less rework (often around 40% less) because bad assumptions get caught in the plan, not in the code.
10-minute action
Pick one non-trivial task you were about to hand to the agent (e.g. “add a priority field to tasks” or “show overdue items on the dashboard”). Do not start in Agent mode. (1) Switch to Ask (or Plan) and describe the goal in 2–4 sentences. (2) Ask for a step-by-step plan and a list of files that would change. (3) Read the plan and fix anything that looks wrong. (4) Only then switch to Agent (or approve the plan) and run implementation. Compare how much you have to undo or fix versus your usual “one shot” run.
What to do next
Try Plan mode on a real task, use the question formats from Day 10 (What if, When should, How would, Why do we) to dig deeper before you approve the plan, and share how it went in the comments.