20 Context Engineering Patterns for Production Systems

The practical, reusable techniques that appear again and again in real context systems. Each is small; together they form a complete working toolkit you can apply to any AI system you build.

Theory is necessary but insufficient. What turns understanding into capability is a repertoire of concrete patterns you can reach for when building. These twenty patterns appear repeatedly across production context systems. None is complicated on its own; mastery comes from knowing which to apply when, and combining them fluently.

Write Patterns

The Scratchpad: have the agent write intermediate findings to external storage, keeping only a summary in active context. The todo.md Plan: maintain a task plan the agent rewrites each step, keeping it in the high-attention zone. Memory Tiers: separate short-term session memory from long-term cross-session memory with different storage. Validate Before Persist: check agent-generated content before writing it to lasting memory, preventing poisoning.

Select Patterns

Just-in-Time Loading: store lightweight references and fetch full data only when the step needs it. Retrieve Then Rerank: pull a broad candidate set semantically, then rerank and keep only the top few. Curated Schema: design a targeted schema of exactly the fields a task needs rather than dumping full documents. Relevance Filtering: score retrieved content and drop anything below a threshold, even if it fits.

Compress Patterns

Threshold Compaction: trigger summarisation automatically when context reaches a utilisation threshold like 75%. Structured Summaries: compress into a defined schema (decisions, constraints, facts) rather than free-form prose. History Windowing: keep only the last N turns verbatim and summarise everything older into a running digest. Context Budgeting: assign a token budget per component and enforce it, cutting by priority.

Isolate Patterns

Sub-Agent Isolation: spawn focused sub-agents with clean contexts for sub-tasks, returning only results. Sandboxed Execution: run tools and code in isolation so verbose output stays out of the main context. Source Reconciliation: when sources conflict, resolve contradictions deliberately before they enter context together. Tool Pruning: include only the tool definitions the current task needs, not every available tool.

Placement and Hygiene Patterns

Recency Placement: put the most critical instructions at the end where attention is highest. Anchor Placement: put stable essential instructions at the very start. Procedural Memory Files: use persistent files like CLAUDE.md for conventions that should always apply. Context Auditing: regularly log and inspect what actually lands in context to find hidden noise. Combine these patterns as your system needs; the best systems chain several together, and fluency means reaching for the right combination instinctively.