Context Engineering

Why This Matters

People often believe AI performance depends only on the model. In reality, even the best model performs poorly when given poor context.

Consider these two requests: "Write a report." versus "You are a project manager. Write a two-page sprint review using the attached Jira issues, meeting notes and company template. The audience is senior leadership."

The second request gives the AI far more context. The model has not changed. Only the information provided has changed. Yet the quality of the output improves dramatically.

This is why Context Engineering has become one of the most important skills in modern AI engineering.

From Prompt Engineering to Context Engineering

Prompt Engineering focuses on writing better instructions.

Context Engineering asks a bigger question: "What information should the AI have before it starts thinking?"

The prompt is only one part of the context. Modern AI systems combine many sources of information before generating a response.

Everyday Analogy

Imagine asking an architect to design a house.

Request 1: "Design a house."

Request 2: "Design a four-bedroom energy-efficient house for a family of five on a 600m² corner lot in a rainy climate with a budget of €450,000."

The second request includes much richer context. The architect makes better decisions because they understand the situation. AI works the same way.

What Is Context?

Context is all the information available to the AI before it generates a response.

Context may include the user request, previous conversation, retrieved documents, uploaded files, company policies, database records, tool results, images, memory, user preferences and system instructions.

The richer and more relevant the context, the better the AI can reason.

The Context Pyramid

Modern AI systems often build context from several layers.

Layer 1 — System Instructions: Who is the AI? Examples include Teacher, Software Engineer, Scrum Master and Legal Assistant.

Layer 2 — User Request: What does the user want?

Layer 3 — Conversation History: Previous questions and answers that maintain continuity.

Layer 4 — Retrieved Knowledge: Information retrieved through RAG such as documentation, wikis, policies and PDFs.

Layer 5 — External Tools: Results from databases, APIs, MCP Servers, search and calendars.

Layer 6 — Memory: User preferences, previous decisions and long-term context.

Together these layers create the complete context presented to the AI model.

Context Window

Every model has a limit to how much information it can process at one time. This limit is called the context window.

Think of it as the AI's working memory. When the context becomes too large, older information may be removed, important details may be lost and performance may decrease.

Good Context Engineering ensures the most valuable information is included.

Context Assembly

Before an AI responds, modern applications often perform several steps.

User Request → Retrieve Documents (RAG) → Call Tools → Read Memory → Load System Instructions → Assemble Context → Send to LLM → Generate Response.

The LLM receives one carefully prepared package of information.

Context Engineering vs Prompt Engineering

Prompt Engineering focuses on writing better prompts.

Context Engineering focuses on designing the entire information environment around the AI. This includes prompts, documents, memory, tool results, retrieved knowledge, user profile and business rules.

Prompt Engineering is a subset of Context Engineering.

Real-World Example

Imagine an AI project manager. A user asks: "Prepare tomorrow's sprint review."

Before answering, the system automatically gathers the sprint backlog, completed issues, Jira updates, team velocity, meeting notes, company presentation template, previous sprint review and user preferences.

Only then does the LLM generate the presentation. The quality comes from the assembled context, not from the prompt alone.

Relationship to Previous Lessons

Prompt Engineering → LLM → Function Calling → AI Agents → Embeddings → Vector Database → RAG → MCP → Context Engineering.

Context Engineering combines everything learned so far into one complete AI system.

Real-World Applications

Enterprise AI: Internal knowledge assistants.

Healthcare: Patient summaries using multiple data sources.

Software Engineering: Code generation using repositories, documentation and issue trackers.

Legal: Contract review using legislation and company policies.

Education: Personalized tutoring using learning history.

Project Management: Sprint reporting using Jira, Confluence and meeting notes.

Research: Combining scientific papers with web search and user notes.

Benefits

Context Engineering enables better accuracy, better reasoning, lower hallucination rates, personalized responses, consistent output, more capable AI Agents and better enterprise AI systems.

Limitations

Good Context Engineering requires high-quality data, effective retrieval, good memory management, well-designed tool integrations, security controls and continuous evaluation.

Poor context produces poor results regardless of model size.

Common Misconceptions

A bigger model does not automatically solve poor context. Even state-of-the-art LLMs need relevant information. Context quality is often a larger factor than model size.

Best Practices

Keep context relevant. Remove unnecessary information. Retrieve only the most useful documents. Structure context clearly. Separate instructions from knowledge. Use citations where possible. Continuously evaluate retrieval quality.

Worked Example: A Context Budget That Ships

A production support assistant on a 128K-window model, designed to a 14K working budget (the AI-061 discipline, applied):

Every row is a decision, and every decision is testable: swap top-3 for top-5 chunks and measure answer quality vs cost (AI-022). That measurement loop, not clever prompt wording, is what context engineering actually is day to day.

Try It Yourself

1. Audit a long chat: after 20+ turns with a chatbot, ask it "What did I tell you in my first message?" If it fails, you've found the history-management edge — then start a fresh chat, paste only a 3-line summary of the old one, and notice how well a curated context performs versus a bloated one.
2. Practice selection: take a 10-page document and a question about page 7. Ask once with the full document, once with just page 7. Compare accuracy, speed, and (if using an API) token cost. You just ran the retrieval-vs-stuffing experiment that justifies this whole discipline (AI-016, AI-061).

Common Mistakes

• Treating context as free because the window is big. Every token costs money, latency, and attention quality (AI-061's lost-in-the-middle).
• Adding instructions to fix every incident. System prompts grow into contradictory sediment; prune on a schedule (AI-057).
• Loading every tool "just in case." Irrelevant schemas dilute selection accuracy (AI-018).
• Never measuring. Context changes without evals (AI-022) are guesses with confidence.

Key Takeaways

• Context engineering = deciding what enters the window: instructions, tools, knowledge, history, and what stays out.
• The window is a budget; allocate it in slots and defend each slot's job.
• Selection beats stuffing: curated context outperforms bloated context at lower cost.
• Every context decision is measurable. Evals turn context engineering from art into engineering.
• This discipline is the bridge from prompting (AI-010) to production systems (Level 3, starting next lesson).