Prompt Engineering

After this lesson, you will be able to: Communicate effectively with LLMs using zero-shot, few-shot, chain-of-thought, and structure techniques.

Prompt engineering is the highest-impact AI skill. Tiny prompt changes can 10x output quality. This lesson covers the patterns that work in production.

Prerequisites:Training and Fine-Tuning LLMs

Anatomy of a good prompt

Role, 'You are an expert SQL analyst'. Task, clear instruction. Context, relevant data, constraints. Format, desired output structure. Examples (few-shot), show 2-3 input/output pairs.

Few-shot vs zero-shot

Zero-shot = describe the task. Few-shot = show examples:

tsx

# Zero-shot
Classify the sentiment: 'The food was cold.'

# Few-shot (often dramatically better)
Examples:
"Loved every bite!" → positive
"Burnt and bland." → negative
"It was fine." → neutral

Now classify: "The food was cold."

Production techniques

1
1. Use XML/JSON structure for inputs and outputs.
2
2. Specify what to do AND what not to do.
3
3. Add output schema with examples.
4
4. For long context: put critical instructions FIRST and LAST.
5
5. Iterate, keep an eval set, measure prompt changes.

Anti-patterns

Don't say 'Don't hallucinate', meaningless. Don't pile on adjectives ('be very accurate and thorough'). Don't bury the task at the bottom of a 5000-token prompt. Be specific, structured, exemplified.

←Training and Fine-Tuning LLMs

Back to Artificial Intelligence

Building with AI APIs→