Context Engineering Framework

Purpose

Context engineering is critical for managing the finite resource of LLM context windows. This skill provides systematic approaches to:

• Measure token usage and identify optimization opportunities
• Compress context without losing critical information
• Optimize token budget allocation across workflow stages
• Handoff work between agents with complete state preservation

Token costs scale linearly with context size - a 100k token context costs 10x more than 10k. Poor context management leads to:

• Failed completions when context overflows
• Degraded quality from truncated information
• Increased costs from redundant context
• Lost work when sessions can't resume properly

Quick Start: 4-Step Process

Step 1: Measure Current Context

# Count tokens in your context
from tiktoken import encoding_for_model
enc = encoding_for_model("gpt-4")
tokens = len(enc.encode(context_text))
print(f"Current context: {tokens:,} tokens")

Step 2: Compress Context

# Apply compression techniques
compressed = apply_compression(context_text, {
    'remove_whitespace': True,
    'deduplicate': True,
    'summarize_verbose': True,
    'extract_references': True
})
print(f"Compressed to: {len(enc.encode(compressed)):,} tokens")

Step 3: Optimize Budget Allocation

# Allocate tokens across workflow stages
budget = TokenBudget(total=100_000)
budget.allocate('system_prompt', 2_000)
budget.allocate('working_memory', 30_000)
budget.allocate('reference_docs', 40_000)
budget.allocate('conversation', 28_000)

Step 4: Create Handoff Document

# Generate handoff for next agent/session
handoff = create_handoff({
    'completed': ['task1', 'task2'],
    'current_state': working_memory,
    'next_steps': ['task3', 'task4'],
    'constraints': ['must_preserve_X', 'avoid_Y'],
    'context_summary': compressed_context
})

Core Patterns Overview

Pattern 1: Token Budget Management

Track token usage across all context components. Set alerts before limits. Enforce hard boundaries to prevent overflow.

Pattern 2: Context Compression (Lossless & Lossy)

Remove redundancy without information loss. Apply semantic compression when acceptable. Preserve critical details while reducing verbosity.

Pattern 3: Semantic Chunking

Split context along natural boundaries. Maintain semantic coherence within chunks. Enable selective retrieval of relevant sections.

Pattern 4: Progressive Summarization

Create multi-level summaries for different detail needs. Drill down when specifics required. Maintain high-level overview for navigation.

Pattern 5: Handoff Document Generation

Capture complete state for work continuation. Include decision history and rationale. Provide clear next steps and constraints.

Pattern 6: Context Window Optimization

Balance detail vs breadth based on task needs. Use RAG for large reference sets. Implement sliding windows for long conversations.

Detailed Documentation

• PATTERNS.md - Full implementation details for all 6 patterns
• KNOWLEDGE.md - Theory, research, and advanced techniques
• EXAMPLES.md - Working code for common scenarios
• GOTCHAS.md - Common pitfalls and debugging strategies
• REFERENCE.md - API documentation and performance data

Top 3 Gotchas

1. Information Loss During Compression

Problem: Aggressive compression removes critical details Solution: Always validate compressed output preserves key information. Use tiered compression with fallback to less aggressive methods.

2. Token Counting Mismatches

Problem: Different models use different tokenizers, counts vary Solution: Always use model-specific tokenizer. Add 10% safety margin to budgets.

3. Handoff Document Incompleteness

Problem: Missing context causes next agent/session to fail or repeat work Solution: Use structured handoff templates. Validate all required fields present.

Quick Reference Card

Token Budget Allocation Guide

System Prompt:      2-5% (2-5k tokens)
Working Memory:    20-30% (20-30k tokens)
Reference Docs:    30-50% (30-50k tokens)
Conversation:      20-40% (20-40k tokens)
Safety Buffer:     10% (10k tokens)

Compression Ratios by Technique

Whitespace removal:        5-10% reduction
Deduplication:           10-30% reduction
Reference extraction:     20-40% reduction
Semantic compression:     40-60% reduction
Aggressive summarization: 70-90% reduction

Context Window Sizes (2024)

GPT-4 Turbo:     128k tokens
Claude 3:        200k tokens
Gemini 1.5 Pro:  2M tokens
GPT-3.5:         16k tokens
Most OSS models: 4-32k tokens

Quick Diagnostic Commands

# Check current token usage
print(f"Tokens used: {count_tokens(context)}")

# Find redundancy opportunities
find_duplicates(context)

# Test compression ratio
test_compression(context, method='semantic')

# Validate handoff document
validate_handoff(handoff_doc)

When to Use This Skill

• Managing projects with 50k+ tokens of context
• Coordinating multi-agent workflows requiring state transfer
• Optimizing costs for high-volume LLM usage
• Debugging context overflow errors
• Implementing long-running conversational agents
• Creating checkpoint/resume capabilities
• Building RAG systems with large document sets

Related Skills

• reverse-engineering-toolkit - Analyze existing context usage patterns
• work-forecasting-parallelization - Estimate token requirements for workflows
• agent-builder-framework - Design agents with efficient context usage
• workflow-builder-framework - Orchestrate context across workflow stages

Integration with Agents

Primary User: context-manager

The context-manager agent should be refactored to use this skill instead of embedding context engineering logic directly.

# Before: Embedded in agent
class ContextManager:
    def compress_context(self, text):
        # 500+ lines of compression logic...

# After: Delegate to skill
class ContextManager:
    def compress_context(self, text):
        return skill('context-engineering-framework').compress(text)

Other Agent Integrations

• agent-orchestrator - Manage context budgets across agent pods
• prompt-engineer - Optimize prompts for token efficiency
• long-running-assistant - Handle session continuity
• document-processor - Chunk and compress large documents

Best Practices

DO

• ✅ Measure before optimizing - profile actual token usage
• ✅ Preserve information architecture during compression
• ✅ Test compression with real workloads
• ✅ Version handoff document schemas
• ✅ Monitor token usage in production

DON'T

• ❌ Compress without validating information preservation
• ❌ Assume token counts are consistent across models
• ❌ Hard-code context window limits
• ❌ Skip handoff validation before agent switches
• ❌ Ignore compression impact on downstream tasks

Validation Checklist

Before using compressed context:

• Key information preserved?
• Semantic coherence maintained?
• References still resolvable?
• Compression ratio acceptable?
• Handoff document complete?

Performance Benchmarks

Technique	Compression Ratio	Info Retention	Speed
Whitespace removal	5-10%	100%	<1ms
Deduplication	10-30%	100%	<10ms
Reference extraction	20-40%	95%	<50ms
Semantic compression	40-60%	85%	<200ms
Aggressive summary	70-90%	60%	<500ms

Getting Started

1. Install dependencies:

pip install tiktoken langchain openai anthropic

2. Run token analysis:

from context_engineering import analyze_context
report = analyze_context("your_context.txt")
print(report.summary())

3. Apply compression:

from context_engineering import compress
compressed = compress(context, target_ratio=0.5)

4. Create handoff:

from context_engineering import create_handoff
handoff = create_handoff(state, next_agent='reviewer')

Conclusion

Effective context engineering is the difference between LLM applications that scale and those that fail under load. This framework provides battle-tested patterns for managing context windows, reducing token costs, and ensuring work continuity across agent boundaries.

Master these patterns to build robust, cost-effective, and scalable LLM systems.

---

For implementation details, see PATTERNS.md. For theory and research, see KNOWLEDGE.md.