Prompt Engineering Challenges

Introduction

Ten focused prompt-engineering challenges to sharpen your intuition. Each one ships with a hidden eval set you can ask the model to generate.

Beginner tip: set a baseline, then iterate one variable at a time.

Understanding the topic

Core concepts:

• Zero-shot vs few-shot on a classification task.
• Force JSON output with a schema.
• Chain-of-Thought for a math word problem.
• Tone-control for a customer email.
• Prompt injection defence test.
• Multilingual prompt that doesn't degrade.
• Length-control prompt (exactly 3 bullets).
• Self-consistency vote on 5 runs.
• Reflection prompt that improves a draft.
• Cost-cut: same quality, 50% fewer tokens.

Syntax reference

Visual workflow / architecture:

┌──────────────┐
│   Exercise   │
└──────┬───────┘
       ▼
  Build · Test · Eval
       │
       ▼
  Compare with model answer

Real-world use

These mirror prompt-eng interviews at Notion AI, Linear AI, Glean, Anthropic and OpenAI.

Best practices

• Track every iteration with score + cost.
• Keep losing prompts in a 'graveyard' file.

Common mistakes

• Tweaking prompts without an eval — random walk.

Hands-on exercise

Interview preparation — practice these questions:

• Q1. How do you measure prompt quality?
• Q2. When does few-shot lose to zero-shot?
• Q3. Scenario: prompt fails 5% of the time. Two strategies?