Prompt Engineering Basics

In this lesson: Write better AI prompts — role, context, format, examples and step-by-step instructions — with before/after prompts.

What is prompt engineering?

A prompt is the instruction you give to an AI model. Prompt engineering is writing that instruction so the answer is accurate, in the right format and useful the first time. The model only knows what you tell it plus what it learned in training – it cannot read your mind, your syllabus or your company rules.

The 6 parts of a strong prompt

Bad vs good prompts

Techniques that work

1. Few-shot examples: show 1–3 examples of input → output. The model copies the pattern far better than it follows a description.
2. Ask for steps: for maths, logic or debugging, ask it to work step by step before giving the final answer, and to show the final answer clearly on its own line.
3. Give the source: paste the chapter text, policy or data and say “answer only from this text; if it is not there, say so”.
4. Specify the format: table, bullet points, JSON with named keys, word limits.
5. Split big tasks: outline first, then each section; or extract data first, then analyse it.
6. Iterate: if the answer is off, tell it exactly what to change instead of starting over.

Few-shot example: classify customer messages

Classify each customer message as ORDER, COMPLAINT, ENQUIRY or OTHER.
Reply with only the label.

Message: "Please send 2 kg kaju katli to Raja Park tomorrow"
Label: ORDER

Message: "The ghewar I bought yesterday was stale"
Label: COMPLAINT

Message: "Are you open on Sunday?"
Label: ENQUIRY

Message: "Do you deliver to Mansarovar and what is the charge?"
Label:

Three examples teach the labels, the style (“only the label”) and the boundary between categories. The answer here should be ENQUIRY.

Prompt templates in Python

When you use an AI model inside an app, build prompts with a function so every request has the same structure and you can test it.

def mcq_prompt(topic, grade, count=5, language="English", avoid=None):
    lines = [
        f"You are an experienced CBSE Class {grade} teacher.",
        f"Create {count} multiple-choice questions on: {topic}.",
        f"Language: {language}. Use NCERT terminology.",
        "Each question: 4 options (a-d), exactly one correct.",
        'Return JSON only: [{"q": str, "options": [str, str, str, str], "answer": "a|b|c|d", "why": str}]',
    ]
    if avoid:
        lines.append(f"Do not include questions about: {', '.join(avoid)}.")
    return "\n".join(lines)

print(mcq_prompt("Refraction of light", 10, count=3, avoid=["lenses", "power of lens"]))

Never trust the output blindly: validate it

Models sometimes add extra text, skip a key or give two correct options. If your program uses the answer, check it with code before saving or showing it.

import json

# a reply copied from an AI model – it added a sentence before the JSON
reply = """Sure! Here are your questions:
[{"q": "Light bends when it passes from air to water. This is called?",
  "options": ["Reflection", "Refraction", "Dispersion", "Scattering"], "answer": "b",
  "why": "Change of speed in a new medium bends light."},
 {"q": "The refractive index of water is about?",
  "options": ["1.00", "1.33", "1.50", "2.42"], "answer": "e", "why": "n = c / v"}]"""

def extract_questions(text):
    start, end = text.find("["), text.rfind("]") + 1      # keep only the JSON part
    data = json.loads(text[start:end])
    good, problems = [], []
    for i, item in enumerate(data, 1):
        if set(item) != {"q", "options", "answer", "why"}:
            problems.append(f"Q{i}: wrong keys {sorted(item)}")
        elif len(item["options"]) != 4:
            problems.append(f"Q{i}: needs 4 options")
        elif item["answer"] not in "abcd" or len(item["answer"]) != 1:
            problems.append(f"Q{i}: invalid answer {item['answer']!r}")
        else:
            good.append(item)
    return good, problems

good, problems = extract_questions(reply)
print("valid questions:", len(good))
print("problems:", problems)

The second question has answer “e”, which is impossible – the code catches it so it never reaches students. Ask the model to fix only that question.

Ready-to-use prompts

Safety and honesty rules

• Never paste passwords, OTPs, Aadhaar/PAN numbers, medical or students’ personal data into public AI tools.
• Verify facts, numbers, laws and citations from official sources – AI can invent them.
• For school work, follow your teacher’s rules and mention when you used AI.
• “Prompt injection”: text inside a document can try to give the AI new instructions. In apps, keep your instructions separate and never let documents trigger actions like sending emails.

Practice tasks

1. Rewrite “make notes on photosynthesis” using all 6 parts of a strong prompt.
2. Create a few-shot prompt that converts informal Hinglish messages into formal English.
3. Extend mcq_prompt() with a difficulty parameter and test the printed prompt.
4. Add a check in extract_questions() that rejects duplicate options.

Summary

• Strong prompt = role + task + context + format + examples + limits.
• Few-shot examples and source text give the most reliable answers.
• In apps, build prompts with functions and validate every output with code.
• Protect private data and verify facts.