Skip to content
learnthevibe

modul 4scaling yourself

plan mode: the plan before the typing

The costliest agent failure is a confident, wrong change across ten files. Plan mode makes the agent research and propose first; you approve before anything touches disk.

~2 min read · ~9 min with the exercise · v1.0 · verified: 2026-07-15

the expensive kind of wrong

By module 3 you shipped a product. Module 4 raises your throughput: planning, parallel work, review, automation. It starts with the cheapest upgrade of the four, because the failure it prevents costs the most: the agent confidently implements a wrong approach across ten files, and you notice at file eight. Reading a plan takes two minutes. Unwinding a wrong implementation takes an afternoon.

Plan mode (a permission mode where the agent reads and explores but edits nothing until you approve its written plan) is the built-in answer. The official docs describe it as the mode where "Claude reads files and runs read-only shell commands to explore but doesn't edit your source files". The agent researches, writes a plan, and waits. Your approval is the gate.

the mechanics

Three ways in (docs): press Shift+Tab to cycle modes mid-session, prefix one prompt with /plan, or start the whole session with claude --permission-mode plan. While the mode is active, the agent explores the codebase and drafts an approach; nothing writes to disk. When the plan lands, you pick from the approval options: execute with auto-approval, execute with edit review, or push back and keep planning. Pushing back is the point. A plan you corrected twice beats a plan you rubber-stamped.

The judgment call is when to bother. A three-line fix needs no ceremony. Reach for a planning pass when the task touches auth, shared data models, routing, migrations, or anything where module 3 raised the stakes. One habit that scales: write the trigger into your CLAUDE.md, so requests above a size threshold start with a plan without you asking.

what it looks like in production

This site's beginner-friendly copy redesign started as a plan file: scope, section order, new components, verification steps, all reviewed before one edit ran:

real session · 2026-07-14 · homepage redesign5 steps
  • redesign the homepage copy for first-time visitors (plan mode)
  • the agent explores read-only: pages, i18n strings, design lawread page.tsx + hu.json + DESIGN.md
  • the plan lands as a reviewable document: scope, files, verificationplan: copy + IA + glossary
  • founder feedback: hero stays, glossary gets its own page
  • approval starts the implementation, on an agreed mapplan approved → execute

a real session from this build: ❯ is the typed request, → are the agent's steps, ✓ is the verification.

what you take with you

  • plan mode reads and proposes; nothing touches disk until you approve
  • read the plan as a reviewer: pushing back twice is cheaper than unwinding once
  • write your own trigger into CLAUDE.md: tasks above a threshold start with a plan

exercise

exercise — plan one real change

Pick a task on your project that touches three or more files and run it through plan mode.

0/4 · log in so we can save your progress

changelog

  • 2026-07-15 · first published, claims validated against the official permission-modes and common-workflows docs

⌘K

search lessons and terms…