Context Engineering

Context engineering is the discipline of systematically selecting, structuring, and delivering the smallest possible set of high-signal tokens that make a task plausibly solvable for an LLM — reliably, safely, and cost-effectively.

This skill provides actionable patterns, checklists, and templates for the full context pipeline — from first principles through production readiness.

When to Use

• Designing context pipelines for AI agents (what enters the LLM context window, in what order)
• Reviewing context quality and token efficiency of existing systems
• Building RAG systems (retrieval pipeline, chunking, packing, grounding)
• Implementing memory (session, working, long-term, organizational, tool-output)
• Designing agent harness (init/continuation prompts, state management, token budgets, caching)
• Multi-agent orchestration (context boundaries, payload schemas, fan-out merge)
• Production readiness assessment for AI systems (8-checklist gate)
• Privacy and compliance for AI features (multi-tenant isolation, consent, data residency)

When NOT to Use

• Prompt technique selection and template patterns → use prompt-engineering skill
• Security audit of individual prompts → use prompt-engineering skill → security-checklist.md
• Non-AI software engineering tasks → use Agent(software-engineer) + stack-specific role
• Infrastructure and deployment → use Agent(devops-engineer)

Context Engineering vs Prompt Engineering

These are complementary disciplines:

• Prompt engineering = designing the instruction text (technique selection, template patterns, formatting, few-shot)
• Context engineering = designing the full context PIPELINE (what enters the window, in what order, how it's managed across turns, how it degrades gracefully)

Context engineering is the system; prompt engineering is one layer within it.

Key Concepts

First Principles

1. Attention is finite — curate, don't stuff. "More context" is not monotonically better
2. Position effects — critical info at beginning and end ("lost-in-the-middle"). Models use tokens near the start and end more effectively than those buried in the middle
3. High-signal tokens > high-volume tokens — compact, discriminative context beats verbose dumps
4. Separation of concerns — policy, knowledge, examples, output contract in distinct layers. Enables independent swap, A/B testing, failure analysis
5. Context is a pipeline — produced by a system (retrieval, ranking, summarization, memory, policy filters, schema enforcement), not concatenated strings

The Context Stack (8-Layer Architecture)

Ordered layers, top = highest priority. Higher layers override lower on conflict.

Rule: keep policy/instructions (Layers 1-2) separate from knowledge (Layer 5). Mixing them is a root cause of prompt injection and "obedience confusion."

→ Full reference: context-stack-model.md

Cacheable Prefix Design

Layers 1-3 rarely change → group all static content at the beginning of the prompt to enable KV cache reuse. Append dynamic content (Layer 4-8) at the end. Never interleave static and dynamic blocks.

Track cache_prefix_ratio = cached_tokens / total_input_tokens (target > 0.3).

Token Budget Allocation

Treat context window as a finite resource with explicit budget per layer:

Overflow strategy: truncate lowest-priority layers first (L7 → L6 → L5).

Resource Files

Integration

• Follows rules: Agent(prompt-engineer) (prompt-level techniques, security, eval-first quality)
• Used by workflows: /ai-assets (context engineering validation), /feature-dev (AI features), /feature-plan (AI work packages), /code-review (context quality review)
• Companion skill: prompt-engineering skill (technique selection, template patterns, eval, security checklist)
• Collaborates with roles: Agent(solution-architect) (AI system design), Agent(product-manager) (agent contracts), Agent(sre-engineer) (context observability), Agent(software-engineer) (implementation)
• Primary source: guides/context_engineering_guide.md