Context Engineering
The discipline of context
Think of it as a desk
Your context window is like a desk with limited space. You can't pile everything on it — you need to choose what's essential, organize it neatly, and keep the most important items within reach. Context engineering is the skill of managing this desk so the LLM can do its best work.
What is Context Engineering?
Context engineering is the discipline of designing and optimizing everything that goes into an LLM's input — the system prompt, user data, examples, history, and instructions. It's about making every token count.
Definition
The systematic practice of selecting, structuring, and prioritizing information within a model's finite context window to maximize output quality.
Why It Matters
Context is the only thing the model sees. Bad context = bad output, regardless of model quality. A well-engineered context can make a small model outperform a large one.
vs Prompt Engineering
Prompt engineering focuses on writing good instructions. Context engineering is broader — it includes what data to include, how to structure it, what to leave out, and how to manage the token budget.
Core Production Skill
In production systems, context engineering determines cost, latency, and quality. It's the difference between a $0.01 API call and a $0.50 one for the same task.
The 5 Pillars of Context Engineering
Every context engineering decision falls into one of these five areas.
Selection — What to Include
Choose the most relevant information. Not everything is useful — including irrelevant data adds noise and wastes tokens. Use relevance scoring, filtering, and RAG to select wisely.
Structure — How to Organize
Order matters. System prompt → instructions → context → examples → user input → output format. Use delimiters (XML tags, markdown) to separate sections clearly.
Compression — How to Fit More
When data exceeds the window, compress it: summarize long texts, chunk documents for RAG, use sliding windows for chat history, or extract key facts only.
Prioritization — What Matters Most
When you can't fit everything, prioritize: current request > recent context > relevant data > examples > old history. Recency and relevance beat completeness.
Budgeting — Token Allocation
Plan your token budget: how much for system prompt, examples, user data, and output reserve. Always leave 20%+ for the model's response.
Common Pitfall: Context Stuffing Without Strategy
The most common mistake is dumping all available information into the context without thinking. This leads to: hitting token limits, drowning the real signal in noise, and paying more for worse results. Always ask: 'Does the model need this to answer the question?'
Getting Started
Audit your current prompts
Count tokens in each section of your prompt. Identify what's essential vs. nice-to-have. Remove redundant information.
Set a token budget
Allocate tokens by priority: system prompt (5-10%), examples (10-20%), user data (40-60%), output reserve (20-30%). Adjust based on task.
Choose a strategy per data type
Small data → stuff it. Large corpus → RAG. Long conversations → sliding window + summarization. Documents → chunking + retrieval.
Measure and iterate
Track output quality vs. context size. Often, a focused 2K-token context outperforms a noisy 100K one. Less can be more.
Fun Fact
Anthropic's research shows that LLMs pay most attention to the beginning and end of the context (the 'lost in the middle' effect). Critical information should go at the start or end of your context, not buried in the middle.
Context Engineering
Try it yourself
Interactive demo of this technique
Write a welcome email to a customer — with and without a system prompt
Dear customer! Welcome to our store. We offer a wide range of products. We hope you enjoy shopping with us. Best regards, Store Team.
Hey there! We're so happy you joined us! You're now part of the SuperShop family, and shopping just got a whole lot better. Here's a little welcome gift — use promo code WELCOME10 for 10% off your first order. Don't be a stranger — reach out anytime! Warmly, Lena from SuperShop
The system prompt is the "DNA" of the response. It sets personality, tone, and rules, transforming boilerplate into a branded message.
Lesson Quiz
1.What is the "lost in the middle" effect in LLMs?
Create a free account to solve challenges
4 AI-verified challenges for this lesson
Related recipes
- Agent Memory: Context That Survives a Restart
- Skills library: SKILL.md for the agent (Phase 6)
- Agent workflow: worktrees and subagents (Phase 7)
- AGENTS.md and ADR: the agent contract for every session
- Architectural barriers for the agent (Phase 3)
- Four modes of developing with AI: A / B / B+ / C
- ExecPlan: an agent on 7–25-hour tasks (Phase 2)
- Harness Configurator: stack → tools and DoD
- Harness Engineering: an environment for an autonomous agent
- Harness garbage collection: drift checks (Phase 5)
- Golden path: a harness template repo (Phase 8)
- Lessons-ledger: Corrections as Code (Phase 5.5)
- LLM-Wiki Second Brain: The Alternative to RAG for Personal Knowledge
- Auditing a brownfield project before the harness (Phase 0)
- Prompt Caching: How to Cut 80% Off Your LLM Bill
This lesson is part of a structured LLM course.
My Learning Path