Instruction-Based Prompting

Introduction

Instruction-based prompts focus on imperative verbs and explicit steps. They're the backbone of every production prompt — clear, testable, edit-friendly.

Beginner analogy: Like writing a recipe vs telling a story. Instructions are: 'preheat oven to 180°C; chop onion finely; cook 5 minutes'.

Understanding the topic

Core concepts to understand:

• Use imperative verbs: 'classify', 'extract', 'rewrite'.
• Number steps when order matters.
• Each instruction should be testable in isolation.
• Avoid stacking too many instructions in one prompt — split or chain.

Syntax reference

Visual workflow / architecture:

BAD:  "Look at this and figure something out."
GOOD: "1. Extract the customer name.
       2. Classify intent (refund | upgrade | other).
       3. Return JSON {name, intent}."

Real-world use

Most production extraction and classification prompts are pure instruction prompts — predictable, debuggable, easy to evaluate.

Best practices

• One instruction = one verifiable behaviour.
• If instructions exceed 7 steps, split into multiple LLM calls.
• Always end with the output format instruction.

Hands-on exercise

Interview preparation — practice these questions:

• Q1. Why are imperative verbs better than passive?
• Q2. How many instructions per prompt is too many?
• Q3. How do you debug a multi-instruction failure?
• Q4. Instructions vs examples — when to use each?