📊 View Lecture Slides Full-screen presentation with navigation

Session 4: Context Engineering

Session Duration: 2 Hours     Block: 1 — Foundations & The Exocortex

Learning Objectives

By the end of this session, students will be able to:

  • Explain the concept of context engineering and how it differs from prompt engineering
  • Design reusable templates for structuring knowledge before AI ingestion
  • Apply formatting techniques to transform raw web content into AI-ready Markdown
  • Identify context design choices that reduce hallucination and improve relevance

Hour 1: Designing What the AI Sees (Instructor-Led — 60 minutes)

1.1 Prompt Engineering vs. Context Engineering

Prompt engineering focuses on how you instruct the model — the words you use to direct its behaviour.

Context engineering focuses on the information you provide alongside those instructions — what you feed the model to inform its response.

Both matter. But context engineering has a more direct impact on whether an AI application produces factually grounded, reliable outputs.

The fundamental principle: Garbage in, garbage out — at scale.

If your knowledge base contains poorly structured, redundant, or contradictory information, your AI application will reflect that. If your context is clean, well-structured, and relevant, the AI has the best possible chance of producing accurate, useful responses.

1.2 The Context Window as a Constraint

Every AI model has a context window — the maximum amount of text it can process in a single call. Current models range from 4,000 tokens (older models) to over 1,000,000 tokens (Gemini 1.5 Pro).

One token ≈ 0.75 words in English.

Tokens Approx. Word Count Practical Equivalent
4,000 ~3,000 words A short essay
32,000 ~24,000 words A short report
128,000 ~96,000 words A small book
1,000,000 ~750,000 words A large textbook

For RAG systems, this means you cannot simply dump your entire knowledge base into every prompt. You must:

  1. Retrieve only the relevant sections
  2. Format those sections efficiently
  3. Fit them alongside the user’s query and system instructions

This is why context engineering matters before RAG is even built.

1.3 Principles of Well-Engineered Context

Principle 1: Remove Redundancy Duplicate information wastes context tokens and can confuse the model. If the same fact appears in three notes, consolidate it.

Principle 2: Be Explicit About Structure AI models perform better when sections are clearly labelled. Use headings, bullet points, and explicit labels.

# BAD (implicit structure)
The product costs £12. It comes in red and blue. Delivery is 5 days.

# GOOD (explicit structure)
## Product Details
- Price: £12
- Colours: red, blue
- Delivery: 5 business days

Principle 3: Include Metadata Date, source, and relevance signals help the model weight information correctly.

## [Source: Official API Documentation | Updated: 2024-06]

Principle 4: Chunk Logically Break documents into semantically coherent sections. A chunk should answer a specific type of question. Avoid arbitrary character-length splits.

1.4 The Context Template Pattern

A context template is a reusable Markdown structure that ensures all notes in your knowledge base follow the same format. For a course knowledge assistant:

---
topic: [Topic Name]
type: [concept | procedure | example | definition]
related: [[related-note-1]], [[related-note-2]]
---

## Core Idea
[One-sentence summary of the main point]

## Explanation
[2–4 paragraphs of clear explanation]

## Key Terms
- **Term 1:** Definition
- **Term 2:** Definition

## Examples
[Concrete example or case study]

## Common Mistakes
[What people typically misunderstand]

Every note written in this format contributes high-quality, retrievable context to your AI system.

Hour 2: Practical — Format Web Data for AI Ingestion (60 minutes)

Lab 4.1 — Clean a Raw Web Page

Find a Wikipedia article or documentation page relevant to your project. Paste the raw content into a text editor.

Tasks:

  1. Remove all navigation text, footer content, and metadata that adds no information value
  2. Structure the content using clear Markdown headings
  3. Convert any tables to Markdown table format
  4. Add a frontmatter block with topic, type, and source fields
  5. Ensure each section answers exactly one question

Save the result as a note in your 04-References/ vault folder.

Lab 4.2 — Apply the Context Template

Using the template from section 1.4, create three structured notes for your project domain. These should cover:

  1. A core concept your project depends on
  2. A procedure your project will perform
  3. An example of the problem your project solves

Lab 4.3 — Test Context Quality

Open AI Studio. Paste one of your structured notes as context. Ask a question that should be answerable from that context.

Now compare:

  • Ask the same question without providing context (pure model knowledge)
  • Ask the same question with your structured note as context

Document the differences in your prompts.md file. Which approach gives a more precise, grounded answer? Does the model cite specific details from your note?

Key Takeaways

  • Context engineering designs the information fed to the AI, not just the instructions
  • Context window limits require deliberate selection of relevant content
  • Well-structured Markdown with explicit headings, metadata, and logical chunks improves AI output quality
  • Context templates ensure consistent, high-quality knowledge base entries
  • Testing context quality before building the retrieval pipeline prevents compounded errors later

Further Reading

  • “How to Think About Context Windows” — Anthropic documentation
  • LangChain text splitter documentation (for RAG chunking strategies)
  • Obsidian Templater plugin documentation