Prompt Engineering

Prompt engineering is the practice of designing the instructions, context, and examples you give a language model so that it produces the output you need — reliably, not just once. This course treats prompts as engineering artefacts: you will structure them deliberately, constrain their output with schemas, diagnose their failures, and measure them with evaluations instead of gut feeling.

Last updated: 01 Nov, 2026

What is prompt engineering?

Definition · Prompt engineering

The discipline of designing, testing, and iterating on the input to a language model — instructions, context, examples, and output constraints — to make its behaviour accurate, consistent, and safe for a specific task.

A prompt that works once in a chat window is not the same as a prompt that works ten thousand times inside a product. Real inputs are messy, users are creative, and models change. The prompts that survive production share three properties: they are specific about the task, constrained in their output, and measured against a test set. That is what this course teaches.

Why it still matters as models improve

Models cannot read your mind. Better models follow instructions more faithfully — which makes clear instructions more valuable, not less.
Your context is unique. Your product's tone, rules, edge cases, and data formats exist only in your prompt.
It is the cheapest lever. Changing a prompt costs minutes; fine-tuning costs days and money.
Every AI system contains prompts. RAG pipelines, agents, classifiers, and evaluators are all driven by prompts.

Prompting vs RAG vs fine-tuning

Approach	What it changes	Best for	Cost to try
Prompt engineering	The instructions and examples sent with each request	Format, tone, task definition, reasoning steps	Minutes
RAG (retrieval)	The facts available to the model at request time	Private, fresh, or large knowledge bases	Hours to days
Fine-tuning	The model's weights	Consistent style or format at scale, specialised behaviour, shorter prompts	Days, plus data and training cost

Always start with prompting. Add retrieval when the model lacks knowledge. Consider fine-tuning only when well-engineered prompts and retrieval still fall short, or when you need to cut cost and latency at high volume.

The prompt engineering workflow

Workflow

1. Define the task      what input, what output, what counts as "correct"?
2. Collect examples     20-100 realistic inputs, including hard and weird ones
3. Draft the prompt     role, task, context, format, constraints, examples
4. Run and inspect      read the outputs - every one, at first
5. Evaluate             score outputs automatically against expectations
6. Diagnose failures    group errors by cause
7. Iterate              change ONE thing, re-run the eval, keep what improves the score
8. Version and monitor  store prompts like code; re-evaluate when models change

What you will learn

Lesson	What you learn
Anatomy of a prompt	The seven building blocks of a prompt, delimiters, ordering, and turning vague prompts into precise ones.
System and developer prompts	The instruction hierarchy, what belongs in the system prompt, guardrails, and versioning prompts as code.
Few-shot examples	Zero-, one-, and few-shot prompting; choosing, formatting, and dynamically selecting examples.
Structured outputs	JSON mode, JSON Schema, Pydantic validation, and repair-and-retry loops.
Failure modes and fixes	Hallucination, format drift, injection, sycophancy, truncation — symptoms, causes, and fixes.
Evaluate a prompt	Test sets, metrics, LLM-as-judge, comparing prompt versions, and regression testing.

Prerequisites

Comfortable Python: functions, dicts, lists, JSON (Python for AI Developers).
Understanding of tokens, messages, roles, and temperature (LLM Fundamentals).
An API key for live experiments. Most code in this course runs offline with simulated models so you can learn the mechanics for free.

Terminal

python -m pip install openai pydantic python-dotenv

Key terms

Term	Meaning
Prompt	Everything the model receives for a request: system message, user message, examples, and context.
Instruction	The part of the prompt that tells the model what to do.
Context	Background information or data the model needs to complete the task.
Delimiter	Markers such as XML tags or triple quotes that separate parts of a prompt.
Zero-shot / few-shot	Prompting without examples / with a small number of examples.
Structured output	Output constrained to a machine-readable format, usually JSON matching a schema.
Eval	An automated test that scores model outputs against expectations.
Prompt injection	Malicious instructions hidden in user input or data that try to override your prompt.

Note

Prompting advice that worked for older models sometimes hurts newer ones — for example, shouting rules in capital letters or adding “think step by step” to a reasoning model. This course focuses on principles that hold across models, and on evaluation so you can verify what works for your model.

Course syllabus

Work through the lessons in order. Your progress is saved in this browser as you mark lessons done.

Syllabus