ngxtm Context Engineering
Context engineering curates the smallest high-signal token set for LLM tasks. The goal: maximize reasoning quality while minimizing token usage.
When to Activate
- Designing/debugging agent systems
- Context limits constrain performance
- Optimizing cost/latency
- Building multi-agent coordination
- Implementing memory systems
- Evaluating agent performance
- Developing LLM-powered pipelines
Core Principles
- Context quality > quantity - High-signal tokens beat exhaustive content
- Attention is finite - U-shaped curve favors beginning/end positions
- Progressive disclosure - Load information just-in-time
- Isolation prevents degradation - Partition work across Task agents
- Measure before optimizing - Know your baseline
IMPORTANT:
- Sacrifice grammar for the sake of concision.
- Ensure token efficiency while maintaining high quality.
- Pass these rules to Task agents.
Quick Reference
| Topic | When to Use | Reference |
|---|---|---|
| Fundamentals | Understanding context anatomy, attention mechanics | context-fundamentals.md |
| Degradation | Debugging failures, lost-in-middle, poisoning | context-degradation.md |
| Optimization | Compaction, masking, caching, partitioning | context-optimization.md |
| Compression | Long sessions, summarization strategies | context-compression.md |
| Memory | Cross-session persistence, knowledge graphs | memory-systems.md |
| Multi-Agent | Coordination patterns, context isolation | multi-agent-patterns.md |
| Evaluation | Testing agents, LLM-as-Judge, metrics | evaluation.md |
| Tool Design | Tool consolidation, description engineering | tool-design.md |
| Pipelines | Project development, batch processing | project-development.md |
| Runtime Awareness | Usage limits, context window monitoring | runtime-awareness.md |
Key Metrics
- Token utilization: Warning at 70%, trigger optimization at 80%
- Token variance: Explains 80% of agent performance variance
- Multi-agent cost: ~15x single agent baseline
- Compaction target: 50-70% reduction, <5% quality loss
- Cache hit target: 70%+ for stable workloads
Four-Bucket Strategy
- Write: Save context externally (scratchpads, files)
- Select: Pull only relevant context (retrieval, filtering)
- Compress: Reduce tokens while preserving info (summarization)
- Isolate: Split across Task agents (partitioning)
Anti-Patterns
- Exhaustive context over curated context
- Critical info in middle positions
- No compaction triggers before limits
- Single agent for parallelizable tasks
- Tools without clear descriptions
Guidelines
- Place critical info at beginning/end of context
- Implement compaction at 70-80% utilization
- Use Task agents for context isolation, not role-play
- Design tools with 4-question framework (what, when, inputs, returns)
- Optimize for tokens-per-task, not tokens-per-request
- Validate with probe-based evaluation
- Monitor KV-cache hit rates in production
- Start minimal, add complexity only when proven necessary
Runtime Awareness
The system automatically injects usage awareness via PostToolUse hook:
<usage-awareness>
Claude Usage Limits: 5h=45%, 7d=32%
Context Window Usage: 67%
</usage-awareness>
Thresholds:
- 70%: WARNING - consider optimization/compaction
- 90%: CRITICAL - immediate action needed
Data Sources:
- Usage limits: Anthropic OAuth API (
https://api.anthropic.com/api/oauth/usage) - Context window: Statusline temp file (
/tmp/ck-context-{session_id}.json)
Scripts
- context_analyzer.py - Context health analysis, degradation detection
- compression_evaluator.py - Compression quality evaluation