Prompt Design Patterns

Introduction

Just as software engineering has design patterns, prompt engineering does too. Recognising patterns lets you reach for the right tool without reinventing each prompt.

Beginner analogy: Like cooking — you don't reinvent boiling water for pasta every time. Patterns standardise the boring parts.

Understanding the topic

Core concepts to understand:

• Classify-then-act — first classify intent, then dispatch.
• Plan-then-execute — outline first, then fill.
• Critique-then-revise — generate, self-critique, rewrite.
• Extract-then-summarise — pull entities, then narrate.
• Few-shot + chain-of-thought — examples with reasoning.
• Role-with-audience — expert speaking to specific listener.

Syntax reference

Visual workflow / architecture:

Pattern              When to use
─────────────────────────────────────
classify-then-act    routing / triage
plan-then-execute    long answers, docs
critique-revise      quality-critical text
extract-summarise    docs & reports
few-shot + CoT       reasoning + format
role + audience      tone-sensitive UX

Real-world use

Most production AI features are 2–3 of these patterns composed together (e.g. classify-then-act + role).

Best practices

• Name your patterns internally — improves team communication.
• Document which pattern each prompt uses in the template header.
• Pattern libraries (LangChain Hub, OpenPrompt) accelerate teams.

Hands-on exercise

Interview preparation — practice these questions:

• Q1. Name 3 prompt design patterns.
• Q2. When use classify-then-act?
• Q3. Why is plan-then-execute better than direct generation?
• Q4. How do patterns affect eval design?