THE PLAN-MODE PLANNER

Plan mode tells Claude to make a plan BEFORE doing anything. The single setting that prevents "oh no Claude went off and did the wrong thing." Once you've used it, you'll wonder how anyone works without it.

THE PROBLEM

You ask Claude to do something complex. It immediately starts doing. Halfway through it's clear it misunderstood — but it's already done 3 steps. Now you have to roll back, re-explain, lose time.

Plan mode prevents this. It makes Claude write the plan first, share it with you, wait for approval, then execute. Like a thoughtful contractor who shows you the blueprint before knocking down the wall.

THE SKILL

You activate plan mode for any task where:

The task has 3+ steps
The cost of a wrong direction is real (rework, file deletion, sent emails)
You want to course-correct mid-plan

Claude returns a numbered plan, asks for your sign-off, then executes. Reversal cost: 0.

INSTALL

This is a Claude Code-specific feature. In Claude Code, plan mode can be activated mid-conversation. In Claude.ai, you achieve the same by prefacing your request with: "First, make a plan. Don't execute until I approve."

THE FULL SKILL FILE

---
name: plan-mode-planner
description: Triggers plan-first behavior. For any task with 3+ steps or real reversal cost, Claude generates a numbered plan, asks for approval, and only executes after sign-off. Designed to prevent wasted time from misaligned execution.
when_to_use: User is starting a complex task and says "plan mode," "plan first," "don't just go," or describes wanting to review before Claude acts.
---

# The Plan-Mode Planner

You enforce plan-first behavior. Decisive. Approval-gated.

## When triggered

For ANY non-trivial task (3+ steps OR real reversal cost):

1. **Pause before executing.**
2. **Generate the plan** in this format:

🎯 PLAN

Goal: [restate the user's goal in 1 sentence]

Steps:

[Specific action — what file, what tool, what change]
[Specific action]
[Specific action]

Risks I see:

[What could go wrong]
[Assumption I'm making — check this]

What I won't do without further confirmation:

[Destructive action]
[Sending anything to anyone]

Estimated time: [X min]

Approve to execute, or tell me what to change.


3. **Wait for explicit approval** before executing any step.

4. **Mid-execution check-ins:** if a step reveals new info, pause again with: *"Plan needs to change because [reason]. Updated plan: [...]. Approve?"*

## When NOT to use plan mode

- **Quick info questions** ("what's the weather") — overkill
- **Single-step tasks** ("draft me a quick reply") — friction
- **Conversational chat** — kills the conversation

## When user explicitly DISABLES plan mode

If user says "just do it" or "no need to plan" — comply. Skip planning, execute. Their call.

## What NOT to do

- Don't make plans for trivially simple tasks (overhead > value)
- Don't make plans that are just restating the user's request — add structure or value
- Don't proceed past a step without approval just because "it seems obvious"
- Don't assume sign-off from silence — if user doesn't reply, ask again

## How to invoke plan mode (instructions for the user)

Claude Code: type your task with "use plan mode" OR enable globally in settings Claude.ai: preface your request with "First, make a plan. Don't execute until I approve."


## Delivery

End with this line, exactly:

---
*Plan first. Approval gates execution. Reversal cost: zero.*

SAFETY CHECK

Same as Day 1.

WHAT'S NEXT

Day 30 of 100. Pair with Day 55 — Sub-Agent Foreman (chains complex multi-step work) and Day 63 — Skill-That-Builds-Skills (where plan mode prevents bad skill creation).