Skip to content

Exercise 1: The Agentic Loop & Plan Mode

The quiz-night API has 5 intentional bugs. You will fix them three different ways to see how Claude's execution model works.

Lab: Lab 01 overview

Theory: The Agentic Loop

What it is. Claude Code is an agentic coding assistant. When you give it a task, it does not produce a single answer; it enters a loop: read the prompt, decide which tool to use (read a file, run a command, edit code), observe the result, then decide the next step. This loop repeats until the task is complete or Claude decides it needs your input.

Why it matters. Understanding this loop is essential because it determines how Claude spends your context window. A vague task may cause Claude to read dozens of files; a precise task lets it go straight to work.

How it works. When you press Enter, Claude Code:

  1. Analyzes your request and the current context
  2. Chooses a tool (Read, Write/Edit, Bash, Grep/Glob, etc.)
  3. Executes the tool and reads the observation
  4. Decides the next step: another tool call, or a response to you

Key actions:

  • Esc: interrupt Claude mid-step
  • Esc Esc or /rewind: undo the last step
  • Shift+Tab: cycle through permission modes

Further reading: Quickstart docs

Theory: Plan Mode

What it is. A permission mode where Claude reads your code and writes a plan, but makes no edits and runs no side-effecting commands. You review, ask for changes, then approve.

Why it matters. The worst failure mode is not "it did nothing," it is "it did the wrong thing for 20 minutes before you noticed." Plan Mode is your preflight check.

How to enter:

Press Shift+Tab until the status line shows the plan mode indicator.

claude --permission-mode plan

You can also set it as default:

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

Gotcha. Plan Mode does not prevent Claude from reading sensitive files, only from modifying them.

Further reading: Permission modes docs

Hands-on

Part 1: Default Mode (watch the loop)

  1. Start Claude Code: 
    claude
  2. Give Claude the task: 
    This API has bugs. Fix them so pytest passes.
  3. Watch Claude work. Notice each tool call, it reads files, runs pytest, edits code, and re-runs pytest.
  4. When pytest passes, note how many tool calls it took.

Part 2: Plan Mode (think before doing)

  1. Reset files and clear context: 
    git checkout -- .
    Then run /clear inside Claude (or /exit and relaunch) to start with a clean session.
  2. Enter Plan Mode via Shift+Tab or: 
    claude --permission-mode plan
  3. Same task: This API has bugs. Fix them so pytest passes.
  4. Read the plan.
  5. Modify it, e.g., "Start with the route handler bug instead."
  6. Switch back to default mode (Shift+Tab) and tell Claude to execute.

Part 3: Interrupt & Rewind (your safety net)

  1. Reset files and clear context: 
    git checkout -- .
    Then /clear (or /exit and relaunch).
  2. Default mode, same task.
  3. After 1–2 tool calls, press Esc to interrupt.
  4. Add: "Also add a comment explaining the purpose of every function you touch."
  5. Let Claude continue. Try /rewind to undo the last step.
  6. Let Claude finish.

Checkpoint

  • [x] pytest passes (all 6 tests green)
  • [x] You cycled through permission modes with Shift+Tab
  • [x] You modified a plan in Plan Mode
  • [x] You interrupted with Esc and used /rewind

← Lab 01 overview · Exercise 2: Context management →