What is Prompt Engineering?

Definition

Prompt Engineering is the practice of designing and refining the text instructions (prompts) you give to an LLM to get better, more consistent results.

It’s the difference between asking “write me some code” and giving a structured request with context, examples, and constraints. A well-crafted prompt can make a $20/month model outperform a $200/month model that receives vague instructions.

Prompt engineering isn’t just about wording — it includes techniques like few-shot learning, chain-of-thought reasoning, role prompting, and structured output formatting.

How It Works

Effective prompts typically include several components:

┌─ System Prompt (role, behavior rules, constraints)
│
├─ Context (background info, documents, examples)
│
├─ Task (what you want the model to do)
│
├─ Format (how you want the output structured)
│
└─ Examples (few-shot: input/output pairs)

The model processes all of these together to shape its response.

Why It Matters

Cost savings — Better prompts reduce retries and token waste
Consistency — Structured prompts produce predictable output formats
Quality — The right technique (CoT, few-shot) unlocks capabilities the model has but won’t show with simple prompts
Safety — System prompts define boundaries and prevent misuse

Example

Bad Prompt

Summarize this article.

Good Prompt

You are a tech journalist writing for a developer audience.

Summarize the following article in exactly 3 bullet points.
Each bullet should be one sentence, focusing on practical implications.
Use present tense.

Article:
{article_text}

Few-Shot Prompt (with examples)

from anthropic import Anthropic

client = Anthropic()

response = client.messages.create(
    model="claude-sonnet-4-6",
    max_tokens=256,
    system="Extract structured data from product descriptions.",
    messages=[{
        "role": "user",
        "content": """Extract the product name, price, and category.

Example 1:
Input: "The new AirPods Pro 2 are available for $249 in the audio category"
Output: {"name": "AirPods Pro 2", "price": "$249", "category": "audio"}

Example 2:
Input: "Get the Galaxy S24 Ultra smartphone for just $1,299"
Output: {"name": "Galaxy S24 Ultra", "price": "$1,299", "category": "smartphone"}

Now extract from:
Input: "The RTX 5090 graphics card is priced at $1,999"
Output:"""
    }]
)

print(response.content[0].text)
# → {"name": "RTX 5090", "price": "$1,999", "category": "graphics card"}

Key Techniques

Technique	When to Use	Effect
Few-shot	Formatting, classification	Shows the model exactly what you want
Chain-of-Thought	Math, logic, reasoning	Forces step-by-step thinking
Role prompting	Tone/expertise control	”You are a senior engineer…”
Structured output	Data extraction, APIs	Request JSON, XML, or markdown tables
Constraints	Limiting scope	”In 3 sentences or fewer”

Key Takeaways

Prompt engineering is a skill, not a hack — it’s how you communicate effectively with LLMs
Start with clear instructions, add examples if the output isn’t right
Use chain-of-thought for anything involving reasoning or math
System prompts set the baseline behavior; user prompts give the specific task
Test prompts with edge cases, not just happy paths

Part of the DeepRaft Glossary — AI and ML terms explained for developers.

Prompt Engineering