Claude Certified Architect - FoundationsCCAR-F Learning Guide
Domain 3 · Lesson 3.4 · 5 min read

Plan Mode versus Direct Execution

Exam taskDetermine when to use plan mode vs direct execution

Plan mode lets Claude explore and design before changing anything; use it for large, multi-file or architectural work and go straight to execution for small, clear changes. Use the Explore subagent to keep noisy discovery out of the main context.

What you need to know

Plan mode

In plan mode Claude reads and explores but does not edit files or run changing commands. It produces a plan for you to review. That makes it safe to investigate a codebase and agree on a design before any change is made, which prevents expensive rework.

Choosing

Use plan mode forUse direct execution for
Large-scale changes across many filesA single-file bug fix with a clear stack trace
Several valid approaches to choose betweenAdding one validation check to one function
Architectural decisions (service boundaries, API contracts)A well-understood change with a clear scope
Library or framework migrations (45+ files)Renaming a local variable
Integrations with different infrastructure needsFixing a typo in a comment

The test is uncertainty and blast radius. If you could describe the exact change in a sentence and it touches one place, execute. If the right approach is not yet clear, or a wrong start would mean undoing dozens of files, plan first.

Plan, then execute

The modes combine. A library migration is planned (explore usages, choose a strategy, agree the order) and then executed directly, following the agreed plan.

The Explore subagent

Discovery is noisy: Grep results, file listings, dependency traces. In a multi-phase task, that output can fill the main context before implementation starts. The built-in Explore subagent does the discovery in its own context and returns a summary, so the main conversation reaches the implementation phase with room to work.

In the exam

Many questions present two tasks side by side (a small fix and a large migration) and ask for the right mode for each. Others describe context running out during a long discovery phase, which points to the Explore subagent.

Check your understanding

Question 1

You must (1) fix an off-by-one error at a known line and (2) migrate 45 route handlers from Express to Fastify. What do you use?

Show answer

Answer: C. The fix is clear and local, so execute it. The migration is wide and has several valid strategies, so plan it first.

Question 2

During a long task, discovery output fills the context before implementation begins. What helps most?

Show answer

Answer: D. The Explore subagent keeps verbose discovery in its own context and returns a summary, leaving the main context free for implementation.

Question 3 · select 2

Which TWO tasks call for plan mode? Select TWO.

Show answer

Answer: A, C. Architectural decisions and wide migrations benefit from exploration and agreement before changes. The other two are small, clear and local.

Apply it: Harbor Lane's mode guide

Exercise · about 10 minCase: Harbor Lane Software

Harbor Lane abandoned its Express to Fastify migration halfway. Write a one-page guide to choosing modes, using the case's upcoming work.

  1. Classify each upcoming item: the date library upgrade (120 files), the audit-log feature, and the weekly small bug fixes.
  2. Describe the migration flow from plan to execution in three or four steps.
  3. Say where the Explore subagent fits.
Check your work
Show model answer

1. Classification

WorkModeWhy
Date library upgrade, 120 filesPlan, then executeWide blast radius; strategy (codemod or by hand, order of packages) must be agreed
Audit-log featurePlan, then executeUnfamiliar domain and design choices (what to log, where to store it)
Small known bugsDirect executionClear cause, one place, easy to verify

2. Migration flow

  1. In plan mode, find every usage and group them by pattern.
  2. Choose the strategy and the order (shared first, then web), and agree it.
  3. Exit plan mode and execute one package at a time, running tests after each.
  4. Return to plan mode only if a new pattern appears that the plan did not cover.

3. Explore. Step 1 of the flow: send usage discovery to the Explore subagent so the hundreds of matches stay out of the main session and only the grouped summary returns.

Sources