Prompt Optimization Challenges

Introduction

Take an existing slow / expensive prompt and improve it without losing quality.

Beginner analogy: Like refactoring code for performance.

Understanding the topic

Core concepts to understand:

• Reduce token count by 30% — quality unchanged.
• Cut latency 50% via streaming + smaller model.
• Move to JSON mode without breaking downstream.
• Add prompt caching — measure cost drop.
• Switch to a multi-step workflow — compare end-to-end.

Syntax reference

Visual workflow / architecture:

Baseline → Optimise → Re-eval → Compare

Real-world use

These are real tasks at every AI company. Bring receipts (numbers) to your interview.

Best practices

• Always re-run evals after optimisation.
• Keep before/after metrics.
• Document tradeoffs.

Hands-on exercise

Interview preparation — practice these questions:

• Q1. How cut tokens 30%?
• Q2. Caching strategy?
• Q3. JSON migration — what breaks?
• Q4. Multi-step tradeoff?