Module 02: Core Loop

Previous: Module 01: Installation & Setup

Goal

Understand how Claude Code actually works: the decision loop, context, and how to structure requests effectively.

---

What You'll Learn

• The complete interaction loop (prompt → analysis → decision → action)
• How Claude reads and understands your project
• How context works and why it matters
• Modes: Normal, Plan, and Think modes
• How to structure effective requests

---

The Complete Loop (Deep Dive)

Every interaction with Claude Code follows this sequence:

┌──────────────────────────────────────────────────────────────┐
│ 1. YOU PROMPT                                                │
│    "Fix the bug in auth.js on line 45"                      │
└────────────────────┬─────────────────────────────────────────┘
                     │
                     ▼
┌──────────────────────────────────────────────────────────────┐
│ 2. CLAUDE READS                                              │
│    - Reads auth.js (full file)                              │
│    - Reads related files (auth-test.js, config.js, etc)     │
│    - Understands the error context                          │
│    - Analyzes call sites where auth.js is used              │
└────────────────────┬─────────────────────────────────────────┘
                     │
                     ▼
┌──────────────────────────────────────────────────────────────┐
│ 3. CLAUDE ANALYZES                                           │
│    - Identifies the root cause                              │
│    - Considers side effects                                 │
│    - Plans minimal changes                                  │
│    - Checks for tests                                       │
└────────────────────┬─────────────────────────────────────────┘
                     │
                     ▼
┌──────────────────────────────────────────────────────────────┐
│ 4. CLAUDE DECIDES                                            │
│    Does the change need tests? → Suggest test updates       │
│    Is the change safe? → Proceed or ask for confirmation    │
│    Should multiple files change? → Show full scope           │
└────────────────────┬─────────────────────────────────────────┘
                     │
                     ▼
┌──────────────────────────────────────────────────────────────┐
│ 5. CLAUDE PROPOSES                                           │
│    Shows you:                                               │
│    - Description of changes                                 │
│    - diff view (what's changing)                            │
│    - Reasoning                                              │
└────────────────────┬─────────────────────────────────────────┘
                     │
                     ▼
┌──────────────────────────────────────────────────────────────┐
│ 6. YOU REVIEW                                                │
│    - Read the diff carefully                                │
│    - Ask questions if unclear                               │
│    - Accept or reject the changes                           │
└────────────────────┬─────────────────────────────────────────┘
                     │
                     ▼
┌──────────────────────────────────────────────────────────────┐
│ 7. CHANGES APPLIED                                           │
│    - Files updated on disk                                  │
│    - You can now test/run the code                          │
│    - Next iteration begins                                  │
└──────────────────────────────────────────────────────────────┘

---

How Claude Reads Your Project

Claude doesn't read everything. It's intelligent about scope.

Example: You ask "Fix the login bug"

Claude will:

1. Search for "login" in your codebase
2. Find auth.js, login.js, auth-controller.js
3. Read those files first (full content)
4. Find callers (what calls these files?)
5. Read tests (if they exist)
6. Find related config (environment variables, constants)

Claude won't read:

• node_modules (excluded automatically)
• .git history (too much data)
• Every file in the project (too slow)
• Your entire codebase unless relevant

Pro Tip

Be specific about scope:

• ❌ "Fix the bugs" → Claude has to guess which files
• ✅ "Fix the login bug in auth.js on line 45" → Claude reads exactly what matters

---

Context: The Key Concept

Context is how much of your conversation Claude remembers. It's finite (~200K tokens).

What Uses Context?

Your prompt:           50 tokens
Claude's response:     500 tokens
File reads:            2000 tokens per file
Previous messages:     accumulated tokens

The Context Meter

/status

Shows:

Context: 67%

This means:

• 0-50%: Lots of room remaining, work freely
• 50-70%: Half-used, be mindful
• 70-90%: Getting tight, use /compact
• 90%+: Critical, you must clean up

/compact - Your Safety Valve

When context reaches 70%+, use:

/compact

This:

1. Summarizes previous conversation
2. Discards old messages
3. Frees up ~40% context
4. Keeps you working in the same session

You can compact multiple times in one session.

---

Modes: Normal vs Plan vs Think

Claude Code has three interaction modes.

Normal Mode (Default)

Claude makes changes immediately after analyzing. Use for:

• Simple bug fixes
• Small feature additions
• Refactoring small sections

Flow: Ask → Analyze (1-2 sec) → Propose → Apply

Plan Mode (/plan)

Claude thinks first, proposes a plan, waits for approval before making changes. Use for:

• Complex features
• Risky changes
• System-wide refactors
• When you're unsure about approach

Flow: Ask → Think → Propose plan → You approve → Analyze → Apply

Example:

/plan
Refactor the authentication system to use JWT instead of sessions

Claude responds with a step-by-step plan for you to review.

Think Mode (/think)

Claude shows extended reasoning, thinking through the problem step-by-step. Use for:

• Understanding complex bugs
• Architectural decisions
• Security analysis
• Performance optimization

Flow: Ask → Extended reasoning → Analysis → Proposal

---

Structuring Effective Requests

The Framework: WHAT, WHERE, HOW, VERIFY

Good requests follow this pattern:

Part	Purpose	Example
WHAT	The goal	"Fix the null pointer bug"
WHERE	The scope	"in src/auth/login.js on line 45"
HOW	Constraints	"without changing the API signature"
VERIFY	Expected result	"All existing tests should pass"

Example Good Request

Fix the bug where login fails for emails with + symbols
WHERE: src/controllers/auth.js, line 78 (email validation regex)
HOW: Update the regex to allow + in emails, but keep existing validation otherwise
VERIFY: Existing tests in tests/auth.test.js should pass

Example Poor Request

Fix the bugs

Claude has to ask follow-up questions instead of solving immediately.

---

Session Context

A session is your current conversation with Claude.

Session Facts

• Starts when you run claude
• Ends when you exit or run /clear
• Not saved by default
• Scoped to one project
• Can be managed with /rewind (go back N steps)

Checkpoint Sessions

To save a session (optional):

/checkpoint save "fixed login, added tests"

Later, restore:

/checkpoint load "fixed login, added tests"

---

Exercise: The Complete Loop

Task: Create a simple utility function

1. Ask with WHAT/WHERE/HOW/VERIFY:

Create a utility function to validate email addresses
WHERE: in src/utils/validators.js
HOW: export as validateEmail(email), return boolean, handle edge cases
VERIFY: Write tests in tests/validators.test.js

2. Claude analyzes:

   • Finds src/utils/
   • Reads existing validators
   • Checks tests directory structure
   • Proposes solution

3. Review the diff:

   • Check the function signature
   • Check the validation logic
   • Review the test cases

4. Accept or iterate:

   Looks good, but make the regex more permissive for + symbols

5. Claude updates and you review again

6. Done: You have a tested, working function

---

Key Takeaways

✓ Every request follows: read → analyze → decide → propose → apply

✓ Be specific (WHAT/WHERE/HOW/VERIFY) for faster results

✓ Context is finite, watch your percentage and /compact at 70%+

✓ /plan for risky changes, normal mode for safe ones

✓ Sessions are temporary, use /checkpoint to save important work

---

Validation: You're Ready If...

✓ You can explain the 7-step loop to someone else ✓ You understand what "context" means and why it matters ✓ You know the difference between Plan and Normal modes ✓ You've used /plan for a complex task ✓ You can check /status and understand the output

---

What's Next?

Module 03: Memory & Config covers:

• Creating your first CLAUDE.md
• How Claude remembers preferences
• Project vs global settings
• Custom configuration

This will teach you how to make Claude Code remember your style and preferences.

---

Completed Module 02? → Ready for Module 03: Memory & Config