Prompt Engineering

The new syntax of AI. Learn how to structure prompts that reduce ambiguity and hallucination — design principles, the prompt types (zero/few-shot, CoT, role), evaluation and refinement.

13.1 What is Prompt Engineering?

Remember from Lecture 10: LLMs are pattern completers, not fact-knowers. They predict the next token from probability. Humans rely on context; machines rely on explicit instructions — the more ambiguity you remove, the better the output.

13.2 Prompt Design Principles

The anatomy of a good prompt

Key principles

• Be specific — state the language/version, the constraints, the goal.
• Give context — explain who the output is for and why.
• Use delimiters — separate instructions from data with """, --- or tags. This also defends against prompt injection.
• Specify the output format — JSON, Markdown, a list — to constrain the output space.
• Ask for reasoning when the task is complex.

13.3 Prompt Types

Zero-Shot prompting

Asking the model to perform a task with no examples — relying purely on its pre-training. Good for simple, common tasks the model has seen thousands of times.

Classify this tweet's sentiment: "I loved the service!"

One-Shot & Few-Shot prompting

Providing one (one-shot) or several (few-shot) examples inside the prompt to teach the model the desired pattern. Also called in-context learning. It drastically improves accuracy for specific formats.

Tweet: "Worst day ever."  -> Sentiment: Negative
Tweet: "It was okay."     -> Sentiment: Neutral
Tweet: "I loved it!"      -> Sentiment: ?

Chain-of-Thought (CoT) prompting

Role prompting

Assigning the model a persona changes its vocabulary, tone and depth. "Act as a 5-year-old" → simple words. "Act as a Network Engineer" → technical terms (TCP/IP, latency, packets). Often set in the System Prompt.

Act as a Senior React Developer at a top tech company.
Interview me on 'React Hooks'. Ask one question at a time
and wait for my answer before grading it.

The Temperature parameter

• Low temperature (≈0.1, or 0) — precise, deterministic. Best for coding, maths, SQL, factual tasks.
• High temperature (≈0.8) — creative, random. Best for brainstorming, storytelling, marketing copy.

13.4 Evaluating LLM Responses

How do you know a response is good? Check it against these criteria:

• Accuracy / factuality — is it correct? Watch for hallucinations.
• Relevance — does it actually answer the question asked?
• Completeness — does it cover everything required?
• Format compliance — did it follow the requested output format (JSON, list)?
• Coherence & clarity — is it well-structured and readable?
• Safety — no harmful, biased or inappropriate content.

13.5 Prompt Refinement Strategies

How to refine a weak prompt

• Add specificity — language, version, exact requirements.
• Add context — explain the audience and purpose.
• Add examples — switch from zero-shot to few-shot.
• Add constraints — "Do not include markdown", "Maximum 100 words".
• Add a format — specify exact keys for JSON output.
• Add CoT — "think step by step" for reasoning tasks.

response = client.chat.completions.create(
    model="gpt-4o-mini",
    temperature=0,        # low temp -> precise, factual
    messages=[
        {"role": "system",
         "content": "You are a senior Python tutor. Be concise. "
                    "If unsure, say 'I don't know'."},
        {"role": "user",
         "content": "Explain list comprehension in one example."}
    ]
)
print(response.choices[0].message.content)

Prompt types are tested constantly — make sure you can tell them apart.

prompt = """Convert the instruction into an API command.

Instruction: Remove user 5    -> Command: DELETE /users/5
Instruction: Remove user 12   -> Command: DELETE /users/12
Instruction: Remove user 88   -> Command: """

# Two examples teach the pattern; the model completes the third.

response = client.chat.completions.create(
    model="gpt-4o-mini",
    temperature=0,
    messages=[{
        "role": "user",
        "content": ("Roger has 5 balls. He buys 2 cans, "
                    "each with 3 balls. How many in total? "
                    "Let's think step by step.")
    }]
)
print(response.choices[0].message.content)