context-engineering

name	context-engineering
description	Context engineering for building production LLM applications: context window management, degradation patterns, optimization strategies, memory system selection, multi-agent architecture, filesystem context patterns, and tool design principles. Use when building LLM apps, RAG pipelines, AI agents, multi-agent systems, or when designing memory, tool APIs, or context strategies for any language model application.
allowed-tools	Bash, Read, Write, Glob, Grep
argument-hint	describe the LLM system you are building (e.g. "RAG pipeline" or "multi-agent orchestrator")
user-invocable	false
disable-model-invocation	true
model	sonnet
effort	low
compatibility	>=1.0
metadata	{"category":"ai-engineering","tags":["llm","context-window","rag","memory","multi-agent","prompt-engineering","tool-design"],"phase":"build"}

Context Engineering for LLM Applications

Production reference for building reliable, efficient LLM-powered systems. Context engineering is the discipline of curating what enters the model's attention window — not prompt writing, but holistic management of system prompts, tool definitions, retrieved documents, message history, and tool outputs.

---

The Attention Budget

Context is a finite resource with diminishing returns. Every token added depletes an attention budget shared across all content. The engineering goal is the smallest high-signal token set that achieves the desired outcome.

Critical attention pattern — lost-in-the-middle: Information buried in the center of context receives 10-40% lower recall accuracy than content at the beginning or end. Always place critical instructions and key facts at the start or end of context.

Context ordering for KV-cache efficiency (stable → reusable → unique):

1. System prompt (never changes)
2. Tool definitions (rarely change)
3. Retrieved documents / examples (reused across requests)
4. Conversation history + current query (unique per request)

This ordering maximizes cache hits, dramatically reducing cost and latency.

---

5 Degradation Patterns

Pattern	What happens	Mitigation
Lost-in-middle	Center content gets less attention	Put critical info at edges
Context poisoning	Hallucination or error enters context, compounds	Validate tool outputs; restart with clean context
Context distraction	Irrelevant info competes for attention budget	Filter ruthlessly; exclude what's not needed now
Context confusion	Model can't tell which context applies to current task	Segment tasks; clear section boundaries
Context clash	Multiple correct sources contradict each other	Priority rules; version filtering

Trigger for compaction: at 70-80% context utilization. Don't wait for failure.

---

4 Optimization Strategies

1. Compaction

Summarize old context when approaching limits. Preserve: key decisions, file paths, error messages, current task. Never compress the system prompt. Use structured sections to prevent silent loss: