Prompting Mistakes Beginners Make

Introduction

Every beginner makes the same prompt mistakes. Recognising them early saves weeks of debugging and thousands of wasted tokens.

Beginner analogy: Like a new driver — the first mistakes are universal. Coast in neutral, ride the brakes, forget the indicator. Prompts have their own equivalents.

Understanding the topic

Core concepts to understand:

• Too vague — "Make it better" with no criteria.
• Too long — 4000 tokens of context for a one-line task.
• No format — then complaining the output is messy.
• Contradictory rules — "be brief and comprehensive".
• No examples for unusual tasks.
• Trusting first output as ground truth.
• Stuffing examples with bias the model will mimic.
• Ignoring temperature & randomness.
• No evaluation — shipping a prompt nobody tested.
• Hard-coding prompts in business code instead of templates.

Syntax reference

Visual workflow / architecture:

Mistake → Fix

Vague           → Add measurable criteria
Long            → Trim, summarise, RAG
No format       → Specify JSON / markdown
Contradictions  → Choose one constraint
No examples     → Add 2–3 few-shot pairs
Trust output    → Add evaluation step
Hard-coded      → Move to template file

Real-world use

Most 'GPT is dumb' tweets are actually 'my prompt is unclear'. Top AI teams publish prompt style guides specifically to avoid these traps.

Best practices

• Pair-review prompts the way you pair-review code.
• Write the desired output FIRST, then design the prompt to produce it.
• Always have a tiny eval set (20 cases) before shipping a prompt.

Hands-on exercise

Interview preparation — practice these questions:

• Q1. What's the most common prompting mistake you see?
• Q2. How would you fix a vague prompt?
• Q3. Why are contradictory instructions so damaging?
• Q4. How do you stop a prompt from being too long?