Use plan mode for non-trivial work

Plan mode makes the agent research and propose an approach before it edits anything. You approve or redirect, then it executes. For multi-step work it is the cheapest point to catch a wrong approach.

When PromptConduit flags this

Claude Code records the active permission mode (default, acceptEdits, plan, auto, and so on) on every submitted prompt. PromptConduit counts the share of prompts whose mode was plan. That share is your plan-mode adoption rate.

The “Try plan mode for non-trivial work” tip fires when the adoption rate is below 15% and there are at least 5 prompts, in both the editor extension (live session) and the web Trends view (selected period).

Note what the metric measures: prompts submitted while in plan mode. Approving a plan switches the session to an editing mode, so the follow-up prompts after approval don’t count toward the rate. One planned task usually contributes a small number of plan-mode prompts, which is why the threshold is low.

Why it matters

A wrong approach found after the agent has edited ten files is expensive to unwind. You pay for the edits, then for the conversation that reverts them, then for the second attempt. Caught at the plan stage, the same mistake costs one message.

Planning also produces better work. The agent has looked at the relevant code and thought through the whole change before touching anything, so the edits are more coherent and you see the scope upfront.

Planning tends to cut interruptions as well. You redirect the plan while it is still text, not in-flight edits.

What to do

Enter plan mode for the next substantial task. In Claude Code:

claude --permission-mode plan

Or press Shift+Tab mid-session until the status bar shows ⏸ plan mode on. For a single prompt, prefix it with /plan.

Review the plan properly. When Claude presents the plan, you can approve it and let it run, approve it and review each edit manually, or choose No, keep planning and say what to change. Press Ctrl+G to open the plan in your editor and edit it directly before Claude proceeds.

Use it where it pays:

• Changes touching more than one or two files.
• Anything you’re unsure how to scope.
• Anything costly to reverse: migrations, refactors, schema or API changes.

Skip it for trivial edits. A one-line fix or a rename is faster to just do. Plan mode is a tool, not a tax, and the 15% threshold reflects that.

Make it the default for a risky repo. To start every terminal session in a project in plan mode, set this in .claude/settings.json:

{
  "permissions": {
    "defaultMode": "plan"
  }
}

The VS Code extension doesn’t read project settings for the starting mode; set claudeCode.initialPermissionMode to plan in your VS Code user settings instead.

Execute somewhere safe. For a large approved plan, run the execution in a git worktree so your main checkout stays clean while the agent works.

How to check it worked

In the editor extension’s coaching tab, the Plan-mode use cell shows the share of prompts run in plan mode for the live session. It reads as a warning below 10% and as good at 30% or more. The Plan-mode % sparkline under “Your trends” tracks it day by day from local history.

Signed-in users can see the same rate on the web Trends view over 7, 30, or 90 days, compared with the previous period. You should also see the Interruptions rate fall, since agreeing the approach upfront removes most mid-turn corrections. To see what a planned session cost, run promptconduit cost in it, or read how to track Claude Code token usage and cost. Setup for the coaching tab is in /docs/editor-extension.