context-engineering（情境工程）

Context Engineering

Overview

context-engineering provides systematic strategies for optimizing Claude Code context window usage. It helps you monitor token consumption, reduce context load, design context-efficient skills, and apply proven optimization patterns.

Purpose: Maximize Claude Code effectiveness while managing token costs and maintaining conversation quality

The 5 Context Optimization Operations:

1. Monitor Context Usage - Track token consumption, identify heavy consumers
2. Reduce Context Load - Remove stale content, minimize loaded files
3. Optimize Skill Design - Progressive disclosure, efficient reference loading
4. Separate Planning/Execution - Keep execution context clean
5. File-Based Strategies - Externalize large data (95% token savings)

Key Benefits:

• 29-39% performance improvement with context editing strategies
• 95% token savings using file-based approaches for large data
• Sustained quality in multi-turn conversations
• Cost reduction through efficient token usage
• Prevention of context overflow and conversation drift

Context Window Sizes (2025):

• Sonnet 4/4.5: 200k tokens (standard), 500k-1M (beta for Tier 4)
• Auto-compaction: Triggers around 80% usage (~160k for 200k window)

When to Use

Use context-engineering when:

1. Approaching Context Limits - Context usage >60-70%, need to optimize before hitting limits
2. Building Large Skills - Creating skills with extensive documentation, need efficient loading strategies
3. Token Cost Management - Reducing API costs through optimization
4. Multi-Turn Conversations - Maintaining coherence across extended sessions
5. Skill Design Phase - Planning context-efficient architecture from start
6. Performance Optimization - Improving response quality and latency
7. Conversation Quality - Preventing drift and maintaining focus
8. MCP Integration - Managing context from Model Context Protocol servers
9. Large Data Handling - Working with extensive datasets or outputs

Prerequisites

• Understanding of Claude context windows
• Access to context monitoring (Claude Code /context command)
• Familiarity with progressive disclosure pattern
• Skills under development or optimization

Operations

Operation 1: Monitor Context Usage

Purpose: Track token consumption, identify context-heavy elements, and detect optimization opportunities

When to Use This Operation:

• Beginning of optimization effort (baseline measurement)
• During development (continuous monitoring)
• When approaching context limits (>60-70% usage)
• Investigating performance issues

Process:

1. Check Current Context Usage Use /context command in Claude Code to view: - Total tokens used - Percentage of context window - Files loaded - Recent tool calls
2. Identify Heavy Consumers

- Large files loaded (>5,000 tokens each) - Extensive conversation history - Many tool call results - Large CLAUDE.md files

1. Analyze Usage Patterns

- Which files are loaded but rarely referenced? - Are all tool results still relevant? - Is conversation history necessary? - Are there duplicate or redundant contexts?

1. Document Baseline

- Record current token usage - Note context-heavy elements - Identify optimization targets

1. Set Optimization Goals

- Target token reduction (e.g., reduce by 20%) - Performance improvement targets - Quality maintenance requirements

Validation Checklist:

• Current context usage measured (tokens and percentage)
• Context-heavy elements identified (files, history, tool results)
• Baseline documented for comparison
• Optimization targets set
• High-impact optimization opportunities noted

Outputs:

• Current context usage metrics
• List of context-heavy elements
• Baseline measurement
• Optimization targets
• Priority optimization opportunities

Time Estimate: 10-15 minutes

Example:

Context Usage Analysis
======================
Current Usage: 145,000 tokens (72% of 200k window)

Heavy Consumers:
1. CLAUDE.md: 25,000 tokens (17%)
2. Large skill files: 40,000 tokens (28%)
   - planning-architect/SKILL.md: 15,000 tokens
   - development-workflow/common-patterns.md: 12,000 tokens
   - review-multi/scoring-rubric.md: 8,000 tokens
3. Conversation history: 30,000 tokens (21%)
4. Tool call results: 20,000 tokens (14%)

Optimization Opportunities:
- Split large CLAUDE.md (25k → 10k target)
- Use references/ loading instead of full files (40k → 15k)
- Clear old tool results (20k → 5k)

Target: Reduce to ~100k tokens (50% of window, 31% reduction)

Operation 2: Reduce Context Load

Purpose: Remove stale content, minimize loaded files, and reduce token consumption

When to Use This Operation:

• Context usage >70% (approaching limits)
• Performance degradation noticed
• Before major operations (clear space)
• Periodic maintenance (every few hours)

Process:

1. Remove Stale Tool Results

- Identify old tool call results no longer needed - Tool results from exploratory work - Superseded information

1. Minimize File Loading

- Only load files actively needed - Use Grep instead of Read for searching - Load specific sections, not entire files - Unload files when done with them

1. Optimize Conversation History

- Context editing auto-clears stale content - Summarize long conversations if needed - Start fresh session for new major tasks

1. Reduce CLAUDE.md Size

- Keep only essential long-term instructions - Move project-specific details to separate files - Use CLAUDE.local.md for temporary preferences - Target: <5,000 tokens for CLAUDE.md

1. Apply Progressive Loading

- Load overview/index files first - Load detailed references only when needed - Use skill references/ on-demand loading

Validation Checklist:

• Stale tool results cleared
• Only necessary files loaded
• CLAUDE.md size optimized (<5,000 tokens if possible)
• Progressive loading applied where relevant
• Context usage reduced measurably

Outputs:

• Reduced token count
• Cleaner context window
• List of removed/optimized elements
• New context usage measurement

Time Estimate: 15-30 minutes

Example Reduction:

Before Optimization: 145,000 tokens (72%)

Actions Taken:
1. Cleared 50 old tool results: -15,000 tokens
2. Unloaded 3 large files no longer needed: -18,000 tokens
3. Optimized CLAUDE.md (split to CLAUDE.local.md): -12,000 tokens
4. Used references/ loading instead of full files: -25,000 tokens

After Optimization: 75,000 tokens (37%)

Reduction: 70,000 tokens (48% reduction)
Quality Impact: None - relevant context maintained

Operation 3: Optimize Skill Design

Purpose: Design context-efficient skills using progressive disclosure, lazy loading, and token-aware architecture

When to Use This Operation:

• Planning new skills (design for efficiency from start)
• Refactoring existing skills (improve context efficiency)
• Building large/complex skills (manage context proactively)
• Creating skill ecosystems (coordinate context usage)

Process:

1. Apply Progressive Disclosure Token Impact: 70-80% reduction vs monolithic (5k vs 20k+ tokens)

- SKILL.md: Overview + essentials only (<1,200 lines, ~3,000-5,000 tokens) - references/: Detailed guides loaded on-demand (300-600 lines each, ~1,000-2,000 tokens) - scripts/: Automation loaded when needed

1. Design for Lazy Loading

- Separate content into focused reference files - Each reference file covers one topic - Load specific reference, not entire skill - Example: Load references/structure-review-guide.md not entire review-multi

1. Optimize File Sizes

- SKILL.md target: 800-1,200 lines (2,500-4,000 tokens) - Reference files: 300-600 lines (1,000-2,000 tokens each) - Keep files focused and concise

1. Use Token-Efficient Formats

- Tables instead of prose (higher information density) - Lists instead of paragraphs (more scannable) - Code blocks for examples (clear and concise) - Quick Reference sections (high-density lookup)

1. Consider Context Budget

- Estimate token usage for skill - Simple skill: 3,000-8,000 tokens total - Medium skill: 10,000-20,000 tokens total - Complex skill: 25,000-40,000 tokens total - Design within budget

Validation Checklist:

• Progressive disclosure applied (SKILL.md + references/)
• SKILL.md <1,200 lines (~4,000 tokens or less)
• Reference files 300-600 lines each
• Files focused on single topics (lazy loadable)
• Token-efficient formats used (tables, lists, code blocks)
• Estimated total tokens within budget
• Skill can be partially loaded (references/ on-demand)

Outputs:

• Context-efficient skill design
• Token budget estimate
• Progressive disclosure plan
• Reference file organization

Time Estimate: 20-40 minutes (during planning phase)

Example:

Skill Design: api-integration

Token Budget Analysis:
- SKILL.md: 900 lines → ~3,000 tokens
- references/ (3 files):
  - api-guide.md: 400 lines → ~1,300 tokens
  - auth-patterns.md: 350 lines → ~1,200 tokens
  - examples.md: 300 lines → ~1,000 tokens
- README.md: 300 lines → ~1,000 tokens

Total if all loaded: ~7,500 tokens
Typical usage: SKILL.md only → 3,000 tokens (60% savings)
With 1 reference: 3,000 + 1,200 → 4,200 tokens (44% savings)

Progressive Disclosure Impact:
- Monolithic (all in SKILL.md): ~7,500 tokens always loaded
- Progressive (SKILL.md + on-demand refs): 3,000-4,500 tokens typical
- Token Savings: 40-60% depending on usage

Design: ✅ Context-efficient with progressive disclosure

Operation 4: Separate Planning from Execution

Purpose: Keep execution context clean by separating exploratory planning from focused implementation

When to Use This Operation:

• Starting complex development work
• Context getting cluttered with exploration
• Need clean context for implementation
• Before critical/focused work

Process:

1. Planning Phase (Separate Session) Characteristics: High context usage, exploratory, broad

- Broad codebase exploration - Research and pattern discovery - Architecture decisions - Task breakdown - Output: Plan documents, task lists

1. Execution Phase (Fresh Session) Characteristics: Clean context, focused, efficient

- Load plan documents (not exploration history) - Focused implementation - Specific file operations - Minimal context bloat

1. Session Transition

- End planning session when plan complete - Save planning artifacts (plans, task lists, decisions) - Start new session for execution - Load only: plan docs, CLAUDE.md, immediate dependencies

1. Maintain Clean Execution Context

- Don't re-explore during execution - Follow plan, don't re-research - Load files as needed, unload when done - Keep focus on implementation

Validation Checklist:

• Planning and execution in separate sessions (when appropriate)
• Planning artifacts saved and documented
• Execution session starts with clean context
• Only plan docs and essentials loaded for execution
• No re-exploration during execution
• Context stays focused on current task

Outputs:

• Clean execution context
• Focused implementation
• Reduced context bloat
• Better performance and quality

Time Estimate: Planning decision (0-5 min), session management (as needed)

Example:

Planning Session (Context: 150k tokens, 75% usage):
- Explored 20 files for research
- Analyzed patterns across codebase
- Made architecture decisions
- Created detailed plan
- Output: skill-plan.md (comprehensive)

[End session, save plan]

Execution Session (Context: 30k tokens, 15% usage):
- Load: CLAUDE.md + skill-plan.md + development-workflow
- Context: Clean, focused, 30k tokens
- Implementation: Follow plan, build skill
- Load references as needed (not all at once)

Result: 80% context reduction (150k → 30k)
Quality: Higher (clean context, focused work)

Research Finding: *"Separating planning from execution keeps implementation context clean"* - confirmed by 2025 best practices

Operation 5: File-Based Optimization

Purpose: Externalize large data to temporary files for on-demand analysis, achieving 95% token savings

When to Use This Operation:

• Handling large data sets (>5,000 tokens)
• Processing extensive outputs (logs, reports)
• Working with large MCP responses
• Managing generated content

Process:

1. Identify Large Data

- Data >5,000 tokens (typically >10,000 characters) - Repeated reference to same large content - Extensive generated outputs - Large MCP server responses

1. Externalize to Files # Save large data to temp file echo "large data here" > /tmp/large-data.txt Instead of keeping in conversation context
2. Reference File Instead of Content Large dataset saved to: /tmp/analysis-data.json (50,000 tokens) To analyze: Read /tmp/analysis-data.json when needed Token Impact: 50,000 tokens → ~500 tokens (95% reduction)
3. Load On-Demand

- Read file only when specific analysis needed - Process in chunks if necessary - Don't keep full content in context

1. Clean Up Temp Files

- Remove temp files when no longer needed - Don't accumulate unused files

Validation Checklist:

• Large data identified (>5,000 tokens)
• Data externalized to files
• File paths documented (reference in conversation)
• On-demand loading used (not full content in context)
• Token savings measured (before/after)
• Quality maintained (can access data when needed)
• Temp files cleaned up when done

Outputs:

• Externalized data files
• File path references
• Significant token savings (often 90-95%)
• Maintained data accessibility

Time Estimate: 10-20 minutes (setup and management)

Example:

Scenario: Analyzing large log file (100,000 tokens)

Before Optimization:
- Full log in conversation context: 100,000 tokens
- Context usage: 50% just for log data

After Optimization:
1. Save log to /tmp/app-log.txt
2. Reference in context: "Log saved to /tmp/app-log.txt (100k tokens)"
3. Read specific sections when needed:
   - Read first 50 lines for overview
   - Grep for errors
   - Read relevant sections on-demand

Token Usage:
- Before: 100,000 tokens in context
- After: ~500 tokens (file reference) + ~2,000 tokens (specific reads)
- Savings: 97,500 tokens (97.5% reduction)

Quality: Maintained - can still analyze log on-demand
Access: Full log available when needed

Research Finding: *"File-based approach achieves 95% token savings"* - proven in 2025 optimization studies

Best Practices

1. Quality Over Quantity

Practice: Focus on relevant, high-quality context rather than loading everything

Rationale: Every piece should be current, accurate, and directly relevant to task

Application: Before loading file, ask: "Do I need this right now for current task?"

2. Progressive Disclosure Always

Practice: Design all skills with SKILL.md + references/ pattern

Rationale: 70-80% token reduction vs monolithic design

Application: SKILL.md <1,200 lines, details in references/ loaded on-demand

3. Monitor Regularly

Practice: Check context usage periodically, especially in long sessions

Rationale: Auto-compaction triggers at ~80%, but proactive monitoring prevents drift

Application: Check /context every 30-60 minutes in active development

4. File-Based for Large Data

Practice: Externalize data >5,000 tokens to files

Rationale: 95% token savings while maintaining accessibility

Application: Save to /tmp/, reference file path, load on-demand

5. Separate Planning from Execution

Practice: Plan in one session, execute in clean session with plan artifacts

Rationale: Keeps execution context focused, prevents exploratory noise

Application: When planning >1 hour, start fresh session for implementation

6. Clean Context Before Critical Work

Practice: Reduce context load before important/complex operations

Rationale: Clean context improves quality and performance

Application: Before complex implementation, clear unnecessary context

7. Use Appropriate Tools

Practice: Choose tools that minimize context usage

Rationale: Some tools add more context than others

Application:

• Grep instead of Read for searching (doesn't load full file)
• Glob for finding files (doesn't load content)
• Task agents for exploration (separate context)

8. Optimize CLAUDE.md

Practice: Keep CLAUDE.md concise (<5,000 tokens), split if needed

Rationale: CLAUDE.md loaded every session, large files waste context

Application:

• Essential standards in CLAUDE.md (project level)
• Temporary preferences in CLAUDE.local.md
• Specific domain knowledge in separate files loaded as needed

Context Budgets for Skills

Simple Skills (Total: 5,000-10,000 tokens)

Structure:

• SKILL.md only or SKILL.md + 1-2 small references
• No scripts or minimal automation
• ~400-800 lines total

Token Breakdown:

• SKILL.md: 3,000-4,000 tokens
• References (if any): 1,000-2,000 tokens each
• README: 1,000 tokens

Example: format-validator, simple helpers

Medium Skills (Total: 15,000-30,000 tokens)

Structure:

• SKILL.md + 3-5 references + scripts
• ~1,500-3,000 lines total

Token Breakdown:

• SKILL.md: 3,000-5,000 tokens
• References: 4-6 files × 1,500 tokens = 6,000-9,000 tokens
• Scripts: 2-3 files × 1,000 tokens = 2,000-3,000 tokens
• README: 1,000-1,500 tokens

Progressive Loading: Load SKILL.md (3-5k) + specific reference when needed (+1.5k) = 4.5-6.5k typical

Example: prompt-builder, skill-researcher

Complex Skills (Total: 40,000-60,000 tokens)

Structure:

• SKILL.md + 7-10 references + 4+ scripts
• ~4,000-7,000 lines total

Token Breakdown:

• SKILL.md: 4,000-6,000 tokens
• References: 7-10 files × 2,000 tokens = 14,000-20,000 tokens
• Scripts: 4-6 files × 1,500 tokens = 6,000-9,000 tokens
• README: 1,500-2,000 tokens

Progressive Loading: Load SKILL.md (4-6k) + 1-2 references as needed (+2-4k) = 6-10k typical

Example: review-multi, testing-validator

Key: Even complex skills only load 6-10k tokens typically (not full 40-60k)

Common Mistakes

Mistake 1: Loading Everything Upfront

Symptom: Context quickly fills with all skill files

Cause: Reading all files instead of progressive loading

Fix: Load SKILL.md first, load references/ only when needed for specific operations

Prevention: Follow progressive disclosure pattern

Mistake 2: Not Monitoring Context

Symptom: Unexpected context overflow, performance degradation

Cause: No visibility into token usage

Fix: Check /context regularly, monitor usage patterns

Prevention: Check context every 30-60 min in active sessions

Mistake 3: Large Monolithic CLAUDE.md

Symptom: Every session starts with 20k-50k tokens used

Cause: Putting everything in CLAUDE.md

Fix: Split CLAUDE.md - essentials only, separate files for detailed knowledge

Prevention: Keep CLAUDE.md <5,000 tokens, use multiple files

Mistake 4: Keeping Stale Tool Results

Symptom: Context bloated with old exploratory results

Cause: Not clearing old tool calls

Fix: Context editing auto-clears, but can manually manage by starting fresh sessions

Prevention: Fresh session for major transitions (planning → execution)

Mistake 5: Not Using File-Based Strategies

Symptom: Large data sets consuming 30-50% of context

Cause: Keeping large outputs in conversation

Fix: Save to /tmp/ files, reference file path, load on-demand

Prevention: Any data >5,000 tokens → externalize to file

Mistake 6: Ignoring Progressive Disclosure

Symptom: Skills with 2,000+ line SKILL.md files

Cause: Not using references/ for detailed content

Fix: Extract detailed content to references/, keep SKILL.md as overview

Prevention: Design with progressive disclosure from start (use planning-architect)

Quick Reference

Context Window Limits (2025)

Token Savings Strategies

Skill Token Budgets

Context Usage Guidelines

Optimization Decision Tree

Is context >70%?
├─ Yes → Reduce immediately (Operation 2)
│   ├─ Clear stale tool results
│   ├─ Unload unnecessary files
│   └─ Start fresh session if needed
│
└─ No → Preventive optimization
    ├─ Is data >5k tokens? → File-based (Operation 5)
    ├─ Building skills? → Progressive disclosure (Operation 3)
    └─ Long session? → Consider planning/execution split (Operation 4)

Quick Optimization Actions

Immediate (When context >80%):

1. Check usage: /context command
2. Clear old tool results (context editing helps)
3. Start fresh session with essentials only
4. Load plan docs, not exploration history

Preventive (During development):

1. Design skills with progressive disclosure
2. Use file-based for large data (>5k tokens)
3. Monitor context every 30-60 min
4. Separate planning from execution (for complex work)

Common Commands

# Monitor context usage
/context

# Read specific lines (not full file)
Read file_path --limit 50

# Search without loading (uses Grep, more efficient)
Grep "pattern" path/

# Find files without loading content
Glob "*.py" path/

# Externalize large data
Bash: command > /tmp/output.txt
# Then reference: "See /tmp/output.txt for results"

Token Estimation

Quick Estimation:

• 1 line of code/text ≈ 3-4 tokens (average)
• 1,000 lines ≈ 3,000-4,000 tokens
• Dense prose: 3-3.5 tokens/line
• Code with comments: 2.5-3 tokens/line
• Tables/lists: 2-2.5 tokens/line (more efficient)

For More Information

• Context monitoring: references/context-monitoring-guide.md
• Reduction strategies: references/reduction-strategies.md
• Optimization patterns: references/optimization-patterns.md
• Analysis script: scripts/analyze-context-usage.py

context-engineering helps you maximize Claude Code effectiveness through strategic context management, ensuring optimal performance, quality, and cost-efficiency throughout development.