10.1. Prompt Engineering Basics

Author

Kamil Filipek

10.1.1. What is Prompt Engineering?

Prompt engineering is the practice of designing and refining the input given to a large language model (LLM) in order to elicit the best possible output — without changing the model’s weights. It is the primary interface between a user or application and an LLM.

Unlike traditional machine learning, where you improve performance by retraining on more data, prompt engineering improves performance by crafting better instructions. This makes it accessible, fast to iterate, and often surprisingly powerful.

The quality of an LLM’s output depends heavily on:

  • Clarity — is the task unambiguously stated?
  • Context — does the model have the information it needs?
  • Format — is the desired output structure specified?
  • Examples — does the model know what “good” looks like?

10.1.2. Zero-Shot Prompting

In zero-shot prompting, the model receives only a task description — no examples. It relies entirely on knowledge acquired during pre-training.

import anthropic

client = anthropic.Anthropic()

def prompt(system: str, user: str) -> str:
    msg = client.messages.create(
        model="claude-sonnet-4-6",
        max_tokens=512,
        system=system,
        messages=[{"role": "user", "content": user}]
    )
    return msg.content[0].text


# Zero-shot classification
response = prompt(
    system="You are a precise text classifier. Reply with only the label.",
    user="""Classify the sentiment of this review as POSITIVE, NEGATIVE, or NEUTRAL.

Review: "The battery lasts forever but the screen is disappointingly dim."

Label:"""
)
print(f"Zero-shot: {response}")
Zero-shot: MIXED

Zero-shot works well for tasks the model has seen many times during training (sentiment, translation, summarization). It struggles with niche domains, specific output formats, or tasks requiring precise reasoning.

10.1.3. Few-Shot Prompting

Few-shot prompting provides \(k\) input–output examples before the actual query. The model learns the pattern from the examples and applies it to the new input — no gradient updates, no training.

few_shot_prompt = """Classify customer feedback into: BUG, FEATURE_REQUEST, COMPLIMENT, or QUESTION.

Feedback: "The app crashes every time I try to export a PDF."
Category: BUG

Feedback: "It would be great if you could add dark mode."
Category: FEATURE_REQUEST

Feedback: "Your support team resolved my issue in under 10 minutes — fantastic!"
Category: COMPLIMENT

Feedback: "How do I reset my password without access to my old email?"
Category: QUESTION

Feedback: "The search function doesn't return results when I use accented characters."
Category:"""

response = prompt(
    system="You are a customer support classifier. Reply with only the category label.",
    user=few_shot_prompt
)
print(f"Few-shot: {response}")
Few-shot: BUG

Why few-shot works: the examples demonstrate the label vocabulary, the expected reasoning pattern, and the output format simultaneously — far more efficiently than instructions alone.

10.1.4. Chain-of-Thought Prompting

Chain-of-Thought (CoT) prompting instructs the model to reason step by step before producing a final answer. It substantially improves accuracy on tasks that require multi-step reasoning: arithmetic, logic, causal inference, and complex classification.

# Without CoT
response_direct = prompt(
    system="Answer concisely.",
    user="""A train leaves Warsaw at 08:15 and travels at 120 km/h.
Another train leaves Kraków (300 km away) at 09:00 traveling toward Warsaw at 90 km/h.
At what time do they meet?"""
)
print(f"Direct: {response_direct}")

# With CoT
response_cot = prompt(
    system="Solve step by step, then state the final answer clearly.",
    user="""A train leaves Warsaw at 08:15 and travels at 120 km/h.
Another train leaves Kraków (300 km away) at 09:00 traveling toward Warsaw at 90 km/h.
At what time do they meet?"""
)
print(f"\nChain-of-Thought:\n{response_cot}")
Direct: 09:45

Chain-of-Thought:
Step 1: By 09:00, the Warsaw train has been traveling for 45 minutes = 0.75 hours.
        Distance covered = 120 × 0.75 = 90 km.
        Remaining gap at 09:00 = 300 - 90 = 210 km.

Step 2: From 09:00, both trains move toward each other.
        Combined speed = 120 + 90 = 210 km/h.

Step 3: Time to close 210 km at 210 km/h = 210 / 210 = 1 hour.

Step 4: They meet at 09:00 + 1 hour = 10:00.

Final answer: The trains meet at 10:00.
Note

Simply appending “Let’s think step by step” to a prompt is often sufficient to trigger chain-of-thought reasoning, even without explicit step-by-step examples. This is called zero-shot CoT and was demonstrated by Kojima et al. (2022).

10.1.5. System Prompts and Persona

The system prompt sets the model’s overall behavior, role, constraints, and output style for the entire conversation. It is processed before any user message and acts as a persistent instruction layer.

# Persona: formal academic writer
academic_response = prompt(
    system="""You are a senior academic writer specializing in computational linguistics.
Write in formal, precise prose. Use passive voice where appropriate.
Cite concepts accurately. Avoid colloquialisms.""",
    user="Explain what word embeddings are in 3 sentences."
)
print("Academic:\n", academic_response)

print("\n" + "─"*60 + "\n")

# Persona: friendly tutor for beginners
tutor_response = prompt(
    system="""You are a friendly tutor explaining NLP to complete beginners.
Use simple words, everyday analogies, and a conversational tone.
Never use jargon without immediately explaining it.""",
    user="Explain what word embeddings are in 3 sentences."
)
print("Friendly tutor:\n", tutor_response)
Academic:
 Word embeddings are dense, low-dimensional vector representations of lexical
 units, derived through the application of neural language models to large
 textual corpora. Semantic relationships between lexemes are encoded as
 geometric proximity within the embedding space, such that words sharing
 contextual distributions occupy similar regions. This representational
 framework has been demonstrated to capture both syntactic and semantic
 regularities, enabling their application across a broad range of downstream
 natural language processing tasks.

Friendly tutor:
 Imagine every word in the English language gets its own address in a giant
 map — that's basically what word embeddings do. Words with similar meanings
 end up living close to each other on this map, so "cat" and "dog" would be
 neighbours, while "cat" and "democracy" would be far apart. Computers use
 these addresses (which are just lists of numbers) to understand what words
 mean and how they relate to each other.

10.1.6. Structured Output

For programmatic use, instruct the model to return structured data (JSON, CSV, Markdown tables) rather than free text.

import json

structured_response = prompt(
    system="You are a data extraction assistant. Always return valid JSON only. No prose.",
    user="""Extract the following fields from the job posting below.
Return JSON with keys: job_title, company, location, salary_range, required_skills (list), experience_years (int).

Posting:
Senior Machine Learning Engineer at DataFlow Inc., based in Warsaw (hybrid).
Salary: 25,000–35,000 PLN/month. We require 5+ years of experience with Python,
TensorFlow or PyTorch, and strong knowledge of NLP and transformer architectures.
Experience with MLOps tools (MLflow, Kubeflow) is a plus."""
)

job = json.loads(structured_response)
for key, value in job.items():
    print(f"  {key:<20} {value}")
  job_title            Senior Machine Learning Engineer
  company              DataFlow Inc.
  location             Warsaw (hybrid)
  salary_range         25,000–35,000 PLN/month
  required_skills      ['Python', 'TensorFlow', 'PyTorch', 'NLP', 'transformer architectures']
  experience_years     5

10.1.7. Prompt Engineering Best Practices

Practice Why it matters
Be explicit about the output format Models default to prose; specify JSON, list, table, etc.
Give a role / persona Anchors the model’s style and domain knowledge
Separate instructions from content Use delimiters (---, """) to avoid ambiguity
Ask for reasoning before the answer CoT improves accuracy on complex tasks
Iterate and version your prompts Small wording changes can shift quality significantly
Test edge cases Check empty input, very long input, adversarial phrasing
Specify length “In 2–3 sentences” prevents both truncation and verbosity 
# Template demonstrating all best practices together
SYSTEM = """You are an expert NLP researcher and teacher.
Respond in valid JSON only. No additional prose."""

TEMPLATE = """Analyze the text below and return a JSON object with:
- "main_topic": string
- "sentiment": "positive" | "negative" | "neutral" | "mixed"
- "key_entities": list of strings
- "readability": "beginner" | "intermediate" | "advanced"
- "summary": string (max 2 sentences)

---
TEXT:
{text}
---"""

text = """The Federal Reserve raised interest rates by 50 basis points on Wednesday,
citing persistent inflationary pressures. Markets responded with a sharp sell-off
in technology stocks, while bond yields climbed to their highest level since 2007."""

response = prompt(system=SYSTEM, user=TEMPLATE.format(text=text))
result = json.loads(response)
for k, v in result.items():
    print(f"  {k:<18} {v}")
  main_topic         Federal Reserve interest rate decision
  sentiment          negative
  key_entities       ['Federal Reserve', 'bond yields', '2007']
  readability        advanced
  summary            The Federal Reserve raised interest rates by 50 basis
                     points amid inflation concerns, triggering a tech stock
                     sell-off and pushing bond yields to 16-year highs.