Claude Code: Context Management
Last updated: 2026-08-31
Context is Claude Code's "working memory" — managing context well means managing Claude Code's understanding ability and your API bill.
💡 Tip: Claude Sonnet's context window is 200K tokens (~150K words). The larger the project, the more important context management becomes — not all files need to be in "working memory" simultaneously.
📋 Prerequisites: Chapter 9 - CLAUDE.md Usage Guide
1. What You'll Learn
- Context window mechanism
- Token consumption analysis and optimization
- Large project context strategies
- Manual context control methods
- Cost vs quality balance
2. Context Window Mechanism
(1) Context Composition
| Component | Typical Share | Description |
|---|---|---|
| System prompt | 2-3% | Claude Code built-in instructions |
| CLAUDE.md | 1-2% | Project conventions |
| Project files | 40-60% | Read source code files |
| Conversation history | 30-40% | Previous Q&A records |
| Code output | 5-10% | Generated code |
(2) Context Lifecycle
TEXT
📖 Display only
Session start → Load CLAUDE.md + system prompt
→ Read relevant files based on instructions
→ Each turn accumulates history
→ Early content truncated when approaching limit
→ Long sessions may degrade in quality
▶ Example 1: View Context Usage
TEXT
📖 Display only
> /cost
Token usage:
System prompt: 4,230 tokens
CLAUDE.md: 1,850 tokens
Project files: 52,400 tokens (12 files)
Conversation: 28,600 tokens (8 turns)
Code output: 8,920 tokens
─────────────────────────────────
Total: 96,000 / 200,000 tokens (48%)
Estimated cost: $1.82
3. Token Optimization Strategies
(1) Reduce Unnecessary File Reads
| Strategy | Effect | How-To |
|---|---|---|
| Precise instructions | Reduce 30-50% | Specify exact files rather than "look at project" |
| Limit scope | Reduce 20-40% | Define working directory in CLAUDE.md |
| Clear timely | Reduce 10-30% | Use /clear in long sessions |
| Small steps | Reduce 15-25% | One thing at a time |
(2) Instruction Precision Comparison
TEXT
📖 Display only
# High consumption (reads entire project)
> Help me refactor this project
# Medium consumption (limited scope)
> Refactor all services under src/services/
# Low consumption (precise files)
> Refactor src/services/user.service.ts, extract validation logic to separate file
▶ Example 2: Token Consumption Comparison
TEXT
📖 Display only
# Vague instruction: ~80K tokens, $1.60
> Add a user search endpoint to the project
# Precise instruction: ~25K tokens, $0.50 (70% savings)
> In src/routes/user.ts, add GET /api/users/search endpoint,
> using searchUsers method from src/services/user.service.ts,
> following the search endpoint format in src/routes/product.ts
4. Large Project Context Strategies
(1) Partition Working
MARKDOWN
<!-- Limit work scope in CLAUDE.md -->
## Current Work Scope
- Primary: src/api/ directory
- Models: src/models/ directory
- Current task: User permission module
## Not Currently Relevant
- src/frontend/ (independent frontend dev)
- src/migrations/ (DB migration complete)
- docs/ (docs maintained separately)
(2) On-Demand Loading
TEXT
📖 Display only
# Don't load all files at once
> ❌ "Read all route files and refactor"
> ✅ "Read src/routes/user.ts and add pagination support"
# Continue after completion
> "Now read src/routes/order.ts and add same pagination"
(3) Session Segmentation
TEXT
📖 Display only
# Split long tasks into short sessions
# Session 1: Model design
claude "Design user permission DB model, only read prisma/schema.prisma"
# Ctrl+D exit
# Session 2: Middleware
claude "Based on the permission model, implement RBAC middleware"
# Session 3: Route integration
claude "Integrate RBAC middleware into all routes needing permission checks"
5. Manual Context Control
(1) In-Session Commands
| Command | Purpose | Scenario |
|---|---|---|
/clear |
Clear conversation history | Topic change |
/compact |
Compress conversation history | Long session optimization |
/cost |
View Token usage | Cost monitoring |
/files |
View loaded files | Check context |
▶ Example 3: Before/After Optimization
TEXT
📖 Display only
# Before: Vague instruction — 156K tokens ($3.12)
> Fix all test failures
# After: Precise instruction — 32K tokens ($0.64)
> Run npm test, only fix 3 failing tests in auth module,
> don't read frontend files
Savings: 80%
❓ FAQ
Q What if 200K tokens isn't enough?
A Use
/clear to reset conversation, or split into multiple short sessions. 200K is usually enough for single tasks; only long sessions may exceed.Q Does
/compact reduce quality?A Slightly. Compression keeps key info but loses conversation details. Recommend
/clear on topic change rather than /compact.Q How to know which files Claude Code read?
A Claude Code displays file paths for each operation. Also use
/files to view currently loaded files.Q Can I prevent Claude Code from reading certain files?
A Write "don't read xxx directory" in CLAUDE.md, but compliance isn't 100%. Use
.claudeignore for sensitive files.Q What happens when context is nearly full?
A Claude Code may start forgetting earlier conversation content; answer quality degrades. Time to
/clear or start a new session.Q Is large Token consumption variation normal?
A Yes. Simple queries may use 5K tokens, complex refactoring 100K+. Instruction precision is the biggest factor.
📖 Summary
- Context window 200K tokens; project files and conversation history are main consumers
- Precise instructions can reduce 30-70% Token consumption
- Large projects use CLAUDE.md to limit work scope, partition processing
- Short sessions + git checkpoints beat ultra-long sessions
/clear,/compact,/costare the three core context management commands
📝 Exercises
- Basic (⭐): Run
/costin a project, record Token consumption for a simple vs complex task. - Intermediate (⭐⭐): Complete the same task with vague and precise instructions, compare Token consumption and quality.
- Advanced (⭐⭐⭐): Design context management strategy for a 50+ file project, achieving 50% Token reduction.
