# Context Engineering Best Practices Research findings on building self-improving AI assistant systems with persistent knowledge bases, from [[Anthropic]]'s engineering guidance. ## Principle Treat context as a finite resource. The goal is finding "the smallest set of high-signal tokens that maximize the likelihood of your desired outcome." ## Persistent Knowledge Structure ### Memory Outside Context Windows Implement external storage for information that persists beyond individual inference cycles: - **Structured note-taking**: Agents regularly write notes to external storage, then pull relevant information back into context when needed - **File-based systems**: Store information hierarchically with meaningful naming conventions and timestamps that provide discovery signals - **Progressive disclosure**: Let agents incrementally discover relevant context through exploration rather than loading everything upfront ### Context Curation Strategies **Compaction**: Summarize context windows approaching their limits, preserving architectural decisions and critical details while discarding redundant outputs. **Sub-agent architectures**: Delegate focused tasks to specialized agents with clean context windows. Each returns condensed summaries to the coordinating agent. ### Information Organization Structure system prompts into distinct sections using XML tagging or Markdown headers. This delineation helps models navigate available context. ## Tool Design for Efficient Retrieval Tools should be: - Self-contained and unambiguous in their purpose - Token-efficient in their outputs - Minimal in number—bloated toolsets create decision confusion Use "just-in-time" context strategies: maintain lightweight identifiers (file paths, queries, links) and dynamically load data via tools. ## Avoiding Common Pitfalls Don't hardcode complex brittle logic seeking exact behavior—this creates fragility. Provide heuristics to guide behavior rather than exhaustive rule lists. ## Application to /evolve These principles inform how Claude's knowledge base should work: 1. Knowledge files are external storage that survives context resets 2. INDEX.md provides discovery signals for just-in-time loading 3. Domain/topic organization enables progressive disclosure 4. Heuristics (not exhaustive rules) guide behavior ## Source [Effective Context Engineering for AI Agents](https://www.anthropic.com/engineering/effective-context-engineering-for-ai-agents) - Anthropic Engineering Blog