Context Manager

Context Manager

Skill for managing context window and token usage in Claude Code sessions.

Overview

Claude Code has a context budget of approximately 190,000 tokens. This skill helps manage this budget effectively through:

• Token usage analysis and monitoring
• Context-efficient query patterns
• Strategic file loading
• Optimization strategies for large codebases

Quick Start

Analyze Current Token Usage

Run the token analysis script to understand context usage:

python scripts/analyze_tokens.py .

This provides:

• Total token count estimation
• Budget usage percentage
• Largest files by token count
• Optimization recommendations

For Specific Directories or Files

python scripts/analyze_tokens.py /path/to/directory
python scripts/analyze_tokens.py src/large_file.py
python scripts/analyze_tokens.py . --budget 190000

When to Use This Skill

Use this skill when:

• Starting work on a large codebase (>50 files)
• Noticing performance degradation or incomplete responses
• Planning context-heavy operations (multi-file refactoring)
• Setting up a new project for optimal Claude Code usage
• Claude mentions being near context limits

Core Principles

1. Load Only What's Needed

DO:

• Request specific files by name
• Use line ranges for large files: view(path, [start, end])
• Load files sequentially for related changes
• Work on one module/component at a time

DON'T:

• Load entire directories without filtering
• Keep unnecessary files in context
• Load the same file multiple times
• Request vague "show me everything" queries

2. Use Targeted Queries

Good queries:

"Update the timeout value in config.py line 45 to 30 seconds"
"Show me the login function in auth.py"
"Fix the validation bug in user.py lines 120-135"

Poor queries:

"Help me with this project"
"Load all the code"
"Show me everything related to authentication"

3. Monitor and Optimize

File size guidelines:

• Small (<1K tokens): Load freely
• Medium (1-5K tokens): Load when needed
• Large (5-20K tokens): Use line ranges
• Very large (>20K tokens): Split or load sections only

Budget zones:

• Green (<70%): Normal operation
• Yellow (70-85%): Be selective
• Red (>85%): Load essentials only, consider reset

Progressive File Loading

Follow this pattern for efficient context usage:

1. Structure → "Show me the project structure"
2. Module → "Show me files in the auth module"  
3. Load → "Load auth.py"
4. Target → "Show me lines 100-150 where the login logic is"
5. Action → "Update line 120 to add validation"

Working with Large Files

Strategy 1: Line Ranges

Instead of loading entire files:

# Load only relevant section
view("src/api.py", view_range=[100, 200])
view("src/models.py", view_range=[1, 50])

Strategy 2: Progressive Disclosure

1. "What functions are in auth.py?"
2. "Show me just the validate_token function"
3. "Now show me where it's called"

Strategy 3: Targeted Modifications

# Precise changes without loading full file
str_replace(
    "Update timeout configuration",
    "config.py",
    old_str="TIMEOUT = 10",
    new_str="TIMEOUT = 30"
)

Context Budget Management

Understanding Token Distribution

Typical session breakdown:

• System prompts & skills: ~40K tokens (21%)
• Conversation history: ~30K tokens (16%)
• Available for files: ~120K tokens (63%)

Monitoring Usage

Run periodic checks:

# Quick check
python scripts/analyze_tokens.py .

# Detailed analysis with JSON output
python scripts/analyze_tokens.py . --json > token_report.json

Warning Signs

Start a new session or reduce context when:

• Claude asks to see previously loaded files
• Responses become incomplete or generic
• Performance noticeably degrades
• Token usage >85% for multiple messages
• Switching to a completely different feature

Bash Command Optimization

Commands add output to context. Minimize verbose output:

# Heavy output ❌
npm install
pip list
git log

# Optimized ✅  
npm install --silent
pip list --format=freeze | head -n 20
git log --oneline -10

Use output redirection:

# Suppress unnecessary output
command > /dev/null 2>&1
command --quiet
command | head -n 20

Project Setup for Context Efficiency

Step 1: Create Context Guidelines

Add a .context-notes.md to your project:

## Context Management

### Key Files
- `src/api.py` (large - use line ranges)
- `src/config.py` (small - load freely)

### Context Strategy  
- Work on one module at a time
- Exclude test fixtures (context-heavy)
- Load models individually

### Exclude Patterns
- `data/` directory (large datasets)
- `legacy/` directory (old code)

Step 2: Add Exclusion Patterns

Copy the .claudeignore template:

cp assets/claudeignore-template.txt .claudeignore

Edit to add project-specific exclusions.

Step 3: Organize Code

Structure for selective loading:

src/
├── core/       # Core logic (load as needed)
├── api/        # API routes (load by route)
├── models/     # Data models (load individually)
└── utils/      # Utilities (load specific files)

Avoid flat structures with many large files.

Advanced Techniques

Technique 1: Component-Based Loading

Work on one component at a time:

1. Identify the component
2. Load only relevant files
3. Complete the work
4. Move to next component

Technique 2: Reference Documentation

For large reference files, create summaries:

• Link to external documentation
• Create concise internal docs
• Use code comments for context

Technique 3: Context Checkpoints

For long tasks:

1. Summarize progress periodically
2. Start fresh session with minimal context
3. Continue with only essential files loaded

Technique 4: Split Large Files

When a file exceeds 500-1000 lines:

• Split by functionality
• Separate frequently changed from stable code
• Group by dependencies

Example:

Before: api.py (2000 lines, ~15K tokens)

After:
api/
├── routes.py (~2K tokens)
├── handlers.py (~3K tokens)  
├── validation.py (~1.5K tokens)
└── utils.py (~750 tokens)

Reference Documentation

For detailed information, see:

• optimization_strategies.md: Comprehensive context optimization techniques, file loading best practices, and directory structure guidelines
• claude_code_specifics.md: Claude Code-specific patterns, tool usage best practices, and session management strategies
• quick_reference.md: Quick reference cheat sheet with token budget rules, essential commands, and common patterns

Common Patterns

Pattern 1: Bug Fixing

"What file contains the authentication logic?"
"Show me the login function in auth.py"
"Update line 45 to fix the validation check"

Pattern 2: Feature Development

"What's the structure of the API routes?"
"Show me an example route handler"
"Create a new /users route following that pattern"

Pattern 3: Refactoring

"List files in the models/ directory"
"Load models/user.py"
"Refactor to use dataclasses"
[Repeat for each file]

Pattern 4: Code Review

"Review auth.py for security issues (lines 1-100)"
"Now review lines 100-200"
[Continue in sections]

Troubleshooting

Issue: Context Feels Heavy

Solutions:

1. Run python scripts/analyze_tokens.py .
2. Start new session if usage >85%
3. Load only essential files
4. Use line ranges for large files

Issue: Claude Asks for Previously Loaded Files

Solutions:

1. Acknowledge context limitations
2. Start fresh session
3. Load files more selectively
4. Use explicit file references

Issue: Slow Performance

Solutions:

1. Check token usage with analysis script
2. Reduce loaded file count
3. Use view ranges instead of full files
4. Exclude unnecessary directories

Issue: Need to Work on Large File

Solutions:

1. Use line ranges to load sections
2. View structure first, then load relevant parts
3. Make targeted str_replace edits
4. Consider splitting if frequently modified

Best Practices Summary

1. Analyze first: Run token analysis before major work
2. Load selectively: Request specific files, not directories
3. Use line ranges: For files >500 lines
4. Query precisely: Targeted questions get targeted responses
5. Monitor budget: Check usage regularly
6. Reset when needed: Start fresh for new features
7. Optimize structure: Organize code for selective loading
8. Document strategy: Create project context guidelines
9. Exclude appropriately: Use .claudeignore for vendor code
10. Think incrementally: One component at a time

Integration with Development Workflow

Daily Development

• Quick token check at start: python scripts/analyze_tokens.py .
• Work on one module per session
• Use line ranges for large files
• Start fresh session for new features

Code Reviews

• Load files individually
• Review in sections for large files
• Focus on changed files only

Refactoring

• Analyze token usage first
• Plan component-by-component approach
• One file at a time
• Test each file before moving on

Debugging

• Identify relevant files first
• Load only those files
• Use line ranges for context
• Make precise changes

Additional Resources

Run analysis script with options:

# Basic analysis
python scripts/analyze_tokens.py .

# Custom budget
python scripts/analyze_tokens.py . --budget 150000

# JSON output for integration
python scripts/analyze_tokens.py . --json

# Exclude additional patterns
python scripts/analyze_tokens.py . --exclude build dist