03: Plan Mode & Thinking Controls

25 minutes | You need: a real multi-file task from your backlog

How Planning Works

Most AI coding failures aren’t bad code — they’re solving the wrong problem. Plan Mode forces Claude into read-only: it can explore your codebase and reason about approach, but it cannot write files or run shell commands. This means you review the strategy before any code is written.

The workflow: Plan → Review → Execute. Mistakes caught at the spec level cost minutes. Mistakes caught in code cost hours.

Do This

1. See what vague gets you

Enter Plan Mode: press Shift+Tab until you see “Plan Mode” (or type /plan).

Give Claude a vague one-liner:

Users should be able to edit their profiles.

Review the plan. Count the assumptions. What did Claude decide for you that it shouldn’t have?

2. Write a real spec

Now give Claude something it can’t misinterpret:

Add profile editing to the user settings page.

Requirements:
- PUT /api/users/:id endpoint with zod validation
- Only authenticated users can edit their own profile
- Editable fields: display name, bio, avatar URL
- Validate: name max 100 chars, bio max 500 chars, avatar must be valid URL
- Return 400 for validation errors, 403 for unauthorized
- Update the existing UserSettings component to include an edit form

Done when:
- Existing tests still pass
- New endpoint has tests for happy path, validation error, and auth error
- Frontend form has client-side validation matching API rules

Compare the two plans. The second one leaves almost nothing to interpretation.

3. Challenge the plan

Pick something the plan assumes:

What about race conditions if two tabs are open? How should we handle concurrent edits?

Use Ctrl+G to open the plan in your editor for direct editing if you want to restructure it.

4. Execute

Once satisfied, exit Plan Mode (Shift+Tab) and:

This plan looks good. Execute it.

Or switch to Auto-Accept Edits mode (Shift+Tab again) if you trust the plan fully. The pipeline: Plan Mode → Normal → Auto-Accept Edits — increasing autonomy as your confidence grows.

5. Thinking controls

Claude has adjustable reasoning depth — use it:

Control	What it does	When to use
Alt+T	Toggle extended thinking on/off (run /terminal-setup first)	Complex debugging, architecture decisions
/effort low	Less reasoning, faster responses	Simple edits, formatting, quick questions
/effort medium	Default reasoning level	Most coding tasks
/effort high	More reasoning per turn	Hard problems, subtle bugs
ultrathink in prompt	High effort for one turn, then reverts	One-off hard problems without changing session effort

Default to medium effort (/effort medium is the default). Only dial up for problems where Claude’s first attempt isn’t good enough. Thinking costs tokens and time.

6. Know when to skip Plan Mode

Plan Mode is for:

• Multi-file features
• Unfamiliar codebases
• Architectural changes
• Anything where “doing it wrong” costs more than 10 minutes to undo

Skip it for single-file, well-scoped changes you can describe in one sentence.

Why this matters

Clear spec → 1 round of review. Vague spec → 3 rounds of back-and-forth. Plan Mode is not overhead — it’s the fastest path to correct code. And thinking controls let you match Claude’s reasoning depth to the problem’s actual difficulty, instead of paying max cost for simple tasks.

Artifact

A completed Plan Mode workflow on a real task — planned, reviewed, executed, and diffed. Familiarity with thinking controls.

Go Deeper

Playbook M03 — Specs Are Source Code for the spec-first paradigm, specification templates, and why communication is now the core technical skill.