11 - Prompt & Context Engineering

How to get reliable, affordable behaviour from a model without changing its weights: writing prompts, constraining outputs, prompting reasoning models, choosing what goes into the context window, caching it, defending it, and optimizing it automatically - all measured with evals.

Where This Module Fits

Prompting is the first lever you pull - before retrieval, fine-tuning or agents - and the fastest to iterate on. Everything later in the course is built on it: RAG is context assembly for knowledge, and agents are prompts, tools and context managed in a loop. Understanding how LLMs work explains why the techniques work: tokenization explains letter-counting failures, attention explains position effects, and post-training explains why instructions and reasoning work at all.

flowchart LR
    F["01 Fundamentals"] --> T["02-03 Techniques"]
    T --> S["04 Structured outputs"]
    T --> R["05 Reasoning models"]
    S --> C["06 Context engineering"]
    R --> C
    C --> K["07 Caching & cost"]
    K --> P["08 Production & security"]
    P --> O["09 Automated optimization"]
    O --> L["🧪 Lab: prompt evals"]

    style F fill:#d8dfe8,stroke:#b0bac8
    style C fill:#dde4dc,stroke:#b0c4b0
    style P fill:#e8e0d4,stroke:#c8b89a
    style L fill:#ddd8e4,stroke:#b8b0c8

Chapter Map

#	Chapter	You will learn	Time
1	Prompt Fundamentals	Prompt anatomy, chat templates, roles and the instruction hierarchy, tokens, sampling and non-determinism	50 min
2	Core Techniques	Zero-/few-/many-shot, what examples teach, chain-of-thought, ReAct, personas	55 min
3	Advanced Techniques	Self-consistency, Tree of Thoughts, least-to-most, chaining, self-refine and its limits	50 min
4	Structured Outputs	JSON mode vs schemas, constrained decoding (XGrammar, llguidance, Outlines), provider APIs, vLLM	50 min
5	Prompting Reasoning Models	Effort controls, goal-first prompts, reasoning state in tool loops, inverse scaling	40 min
6	Context Engineering	Long-context evidence, write/select/compress/isolate, arranging and budgeting the window	45 min
7	Prompt Caching & Cost	How prefix caches work, provider comparison, cache-friendly design, other cost levers	40 min
8	Prompts in Production	Templates, versioning, eval gates, model drift, prompt injection and layered defences	55 min
9	Automated Prompt Optimization	APE, OPRO, DSPy with MIPROv2 and GEPA, metrics and overfitting	55 min
10	Q&A Review Bank	56 questions across the module	60 min

Code Lab

Lab	What you build	Runs on
Prompt Evals & Structured Outputs	An eval harness comparing zero-/few-shot and free/constrained decoding on a local model, with bootstrap CIs	Laptop CPU, ~2 min

Mini-Project

Build and ship a ticket triage prompt for a real or synthetic dataset of at least 200 labelled items:

1. A versioned prompt template with category definitions and an enforced output schema.
2. An eval harness reporting accuracy per category with confidence intervals, gated in CI on a paired comparison.
3. A cost report: tokens per call, cache hit rate, and cost per 1,000 tickets at two reasoning-effort levels.
4. A security test set of 20 injection attempts embedded in tickets, with the defence layers you applied and their measured success rates.
5. One optimization run (few-shot bootstrap or MIPROv2) evaluated on a held-out test set.

Deliverables: the repository, a one-page results summary with intervals, and a short note on what you would change for production.

Review

• Q&A Review Bank - consolidated questions for this module
• Module quiz - every Check Yourself question in this module, in course order

---

Previous: 10 - Cloud Platforms · Next: 12 - RAG

Section Appendix

Summary & Key Terms - a quick recap of this section and its essential vocabulary.