Prompt Engineering - Prompt Examples

Imagine an auto driver in Hyderabad.

• Say "Take me somewhere" → the driver is confused — which place?
• Say "Take me to Ameerpet Metro Station" → the driver understands and takes you there.

Same driver, same auto — the only difference is a clear instruction. AI works exactly like this.

When you send a prompt, AI follows a simple four-step process:

• You type a prompt — for example, "Write a birthday message for my friend".
• AI reads your instruction — it looks at the words and the task.
• AI understands the meaning — it finds what you actually want.
• AI generates the output — it produces the answer in the style you asked.

A bad prompt is unclear and missing details; a good prompt is clear, specific, and gives proper direction. The quality of your question decides the quality of the answer.

A vague prompt like "Explain Java" gives a general answer because the level, audience, and topic are unknown. A specific prompt tells AI exactly what to cover:

• Topic — what to explain.
• Level — beginner, intermediate, or advanced.
• Audience — who it is for.
• Style — simple, formal, funny, etc.

An incomplete prompt is missing details. "Write resume" gives a generic resume because AI does not know your skills or experience. A complete prompt gives all the information AI needs.

Here is the same idea as a ready-to-use prompt. Notice all five parts appear one by one:

One more example for interview practice:

Tone is the style or feeling of the response. It controls how AI speaks, not what it speaks.

• Same question + different tone → different style of output.
• Like asking for chai in a friendly, polite, strict, or funny way.

Length control means deciding the size of the answer — short, medium, or long. If you do not mention length, AI decides on its own, and the answer can be too long or too short for your need.

Short answers save time and are perfect for revision. Long answers give deep understanding and are perfect for learning a topic properly.

Step-by-step prompting asks AI to explain in ordered steps. Instead of one confusing wall of text, AI gives a structured explanation where each step builds on the previous one.

The real power comes from combining all three controls in one prompt:

Reading about prompting is like reading a driving book — it does not make you a driver. Real skill comes from doing.

• Take a bad prompt.
• Improve it using clarity + formula.
• Test it with AI.
• Refine it again — repeat daily.

Format control means telling AI how the answer should be presented — not what to answer, but how to show it.

• The same content can look different in bullet points, a table, steps, a checklist, or a story.
• Without a defined format, AI gives a messy wall of text.

The restaurant example explains it best:

• "Give food" → confusion.
• "Give veg meals — rice, dal, curry, neatly in a plate" → a perfect serving.

Same food, different presentation. AI works the same way.

Structured output means forcing AI to follow the same fixed format every time — same headings, same order, same structure.

• The output becomes predictable, consistent, and reusable.
• Like an application form with fixed fields instead of a random paper.

The word "strictly" (or "follow this structure strictly") is the key:

• Without it → AI may quietly change the headings or add its own sections.
• With it → the output stays locked to your format.

By default AI talks extra — it adds an introduction, greetings, explanations, and a summary you never asked for. The "Only Output This" control tells AI to give only the required output and nothing else.

• Essential for JSON, apps, APIs, and automation — extra text breaks the data.
• Think of an ATM: you want cash, not messages. Tool output must be pure data.

• Essential for JSON, apps, APIs, and automation — extra text breaks the data.
• Think of an ATM: you want cash, not messages. Tool output must be pure data.

Useful phrases: "Reply only with output", "do not add explanation", "no greetings", "no summary".

A reusable prompt is one fixed prompt that you use for many different topics. It combines format control + structured output + only-output control into a single system.

• One prompt in → unlimited structured notes out.
• Use the exact same prompt for SQL joins, Python lists, operating systems — anything.
• The output always comes back in the same clean structure.

An LLM understands only written instructions, not your intention. It reads your prompt from start to finish, like a command list, and processes each instruction in order.

• If an instruction is missing or confusing, the output changes.
• AI follows written instructions, not your intention.

This is why instruction order matters: put the role first, then the task, then context, then constraints. A prompt that starts with "Act as a teacher" gives far clearer direction than one that mentions the role at the very end.

Prompt engineering is word-level precision. One small word can change the length, tone, or format of the entire output.

• "Bring vegetables" → anything.
• "Bring only onions" → restricted.
• "Bring exactly 1 kg onions" → controlled.

Ambiguity means an unclear or incomplete instruction. When the prompt is not clear, AI fills the missing gaps by guessing based on probability — and the output becomes random. The problem is not the AI; the problem is the unclear instruction.

The friend example explains it:

• Say "Bring it" to a friend → they guess — wrong item, wrong place.
• The problem is not the friend, it is the unclear instruction.

AI behaves exactly the same way.

Constraints are rules that limit AI behavior. Each type controls one thing:

• Length constraint — controls size.
• Tone constraint — controls style.
• Format constraint — controls structure.
• Plain rules — control what AI may or may not do.

Without constraints, AI has too much freedom and the output becomes inconsistent.

An example is a sample output that guides AI's style. When you show AI one good example, it follows the same pattern — this is called giving a reference.

The carpenter story shows it:

• "Make table" → confusion.
• "Make a study table, 4 feet, brown colour" → correct.
• Showing a sample image → perfect.

Zero-shot prompting means asking AI directly without giving any example — no sample, no format example. AI answers purely from its training. It is the simplest form of prompting.

• Shop example: say "Give pen" → the shopkeeper gives you any pen — no example, based on guess.
• AI works the same way in zero-shot mode.

One-shot prompting means giving one example to guide AI. The single example shows AI the expected format, tone, and structure, and AI copies that pattern. One example = strong direction.

• Tailor example: show one shirt → the tailor stitches the same style — no guessing, same pattern.
• Useful for emails, notes, and interview questions where you want a consistent look.

Few-shot prompting means giving 2–5 examples to teach AI a pattern. AI learns from the examples and applies the same logic to new input.

• Few-shot is the teacher who gives three examples — one gives little understanding, three give clear understanding, five give mastery.
• Works best for classification, MCQ generation, email replies, and any task with a clear pattern.
• Keep the examples consistent — mixing different patterns confuses AI.

Example 1 — MCQ generator:

Example 2 — email reply:

Few-shot + structure = production-ready AI output. Each piece does one job:

• Few-shot — teaches the pattern.
• Structure — locks the format (table, JSON, headings).
• Strict rules — remove extra text.

Notice the two instructions working together: the example taught AI the pattern, and the words "in JSON format only" locked the structure.

• Bank statement, invoice, resume — extract any of them into JSON with this technique.

Persona control means controlling how AI behaves — its strictness, its depth of explanation, and its attitude. A persona is not just a role; it is behaviour + depth + attitude.

Notice how the same topic would get very different outputs from these two personas — one challenges, the other encourages. That is persona control in action.

A persona alone gives identity, but to make output stable and professional you combine three layers:

• Persona — behaviour.
• Constraints — rules.
• Format — structure.

This combination is what turns a one-time prompt into a reusable AI tool. The school-exam example shows why:

• "Write about water" → different answers from every student.
• "Write 5 points, each one line" → the same pattern from everyone.

A domain persona makes AI think from a specific field — HR, analyst, mentor, CEO.

• The same problem gets completely different answers depending on the domain.
• Each field focuses on different things.

The cricket example: watch the same match as a fan (emotion), a coach (technique), and a selector (talent) — same event, different perspective.

• Domain persona gives AI the right perspective for the problem.
• Each domain sees the same problem through its own lens.

A safety persona makes AI behave safely — politely refusing harmful, illegal, or cheating requests and suggesting safe alternatives.

• In production systems, safety personas are not optional.
• They protect users and the system itself.

Good safety prompts combine three parts:

• Clear identity — "You are a safe assistant".
• Clear rule — "refuse harmful requests".
• Graceful alternative — "suggest a safe alternative".

The step-by-step instruction forces AI to work through the logic instead of guessing. Use this for aptitude problems, discounts, debugging, decisions, and anything with logic.

Manual Chain of Thought (Manual CoT) means you write the thinking steps explicitly in the prompt. You guide AI step by step, and AI follows your structure — you control the thinking process.

The travel example:

• "Go Hyderabad" → confusion.
• "Go via ORR, take the exit, go straight, turn left" → zero confusion.

Clear steps in the prompt give AI a clear path.

Auto Chain of Thought (Auto-CoT) lets AI generate the step-by-step examples automatically. Instead of you writing every step, AI creates solved examples first, then you reuse that pattern for new problems.

• It saves time.
• It is great for learning.

The driving example:

• Beginner → thinks through every step (manual).
• Expert → drives automatically (auto).

Same driving, less effort. Once you understand the pattern, let AI generate examples and apply them to new problems.

Efficiency means getting the best output with the minimum words and time. Not every question needs step-by-step reasoning:

• Simple questions → quick answers.
• Complex questions → full reasoning.

Balance depth and speed.

Reflection means asking AI to review and improve its own answer. The first answer may not be perfect. Reflection adds a checking step:

Generate → Review → Fix → Give the improved final answer.

• Exam example: you write the answer, check it again, and correct the mistakes — that checking is reflection.
• AI can do the same when you ask it to.

Self-consistency means solving the same problem multiple times and comparing the answers. If AI gives the same answer every time, you can trust it more.

• Three-friends analogy: ask three friends and choose the common answer — that is self-consistency.
• If all three give the same answer, it is probably right.

For important tasks, use both together: solve multiple times, then reflect on the best answer.

Tree of Thoughts means exploring multiple thinking paths and choosing the best one. Instead of one straight answer, AI creates branches — several options — compares them, and picks the winner. Think like a leader, not just a solver.

The road example:

• Three roads are available.
• A smart person checks all three, compares the traffic, and chooses the best — that is ToT.
• AI does the same when you ask it to generate and compare options.

Chain of Hindsight means learning from past mistakes and improving future answers. AI generates an answer, finds its own mistake, learns from it, and gives a corrected version — mistakes become improvements.

• Cricket example: a player gets out, reviews the mistake, and improves the next match — that is hindsight.
• AI can review its own answer the same way.

Step-Back reasoning means moving back and understanding the big concept before solving the problem. Instead of jumping straight into the solution, AI first explains the core idea, then solves.

• Batting example: before batting, a player studies the pitch and conditions — that is step-back.
• It stops you from jumping into a solution without understanding what you are solving.

Notice how the prompt forces AI to identify components, connect them, and analyze impact — not give a single answer. This is system-level thinking used by senior planners and managers.

Reliability engineering means making AI output correct, safe, consistent, and trusted — not random answers, not dangerous outputs, but controlled and verified output. The goal is accuracy + safety + trust.

Two everyday examples:

• Highway example: a car without barriers can have an accident; with guardrails it is safe — AI needs guardrails too.
• Bank example: the cashier counts the money twice — that is validation. AI output needs the same validation.

Guardrails are the allowed / not-allowed rules that limit what AI may answer. A hallucination is when AI creates fake facts with confidence.

• Guardrails → allowed / not-allowed topics.
• Hallucination control → stop AI from creating fake facts.
• Both are controlled with clear prompt rules.

Complex case simulation means applying multiple thinking techniques together. Real problems are too big for a single method:

• CoT — break the problem.
• ToT — compare options.
• Self-Consistency — multiple solutions.
• Reflection — improve.

ReAct means Reason + Act. Thinking alone is not enough — AI should think what is needed, take an action, re-evaluate, and repeat until the goal is complete.

• AI becomes an assistant that completes tasks, not just an advisor that gives ideas.
• The loop repeats: think → act → re-evaluate → repeat until done.

Notice the pattern: every step alternates between thinking (Reason) and doing (Act). This loop is the foundation of AI agents.

Tool calling means AI uses an external helper to get an accurate result. AI should not guess; it should verify.

The student example:

• Calculator for math.
• App for train timing.
• Dictionary for spelling.

Tools reduce mistakes — AI works the same way.

A stop condition is a rule that tells AI when to stop. Without one, AI can loop forever — an infinite loop that wastes time and money.

Stop conditions + retry limits + output limits = controlled execution.

The cleaning-room example:

• Without a rule → you clean forever.
• With the rule "clean for 30 minutes, stop when the room is clean" → you finish.

AI needs the same kind of boundary.

A milestone is a big checkpoint — like "content ready" or "batch launched".

• Milestones let you measure progress.
• They keep the project on time.

This single prompt combines CoT, ToT, CoH, Reflection, ReAct, guardrails, and a stop condition — everything you learned so far, working together as one agent.

The capstone is the final project that combines all your skills. It turns individual topics into a combined system that solves a real problem.

The school-annual-day example:

• You practise for many days.
• Then you perform on the final day — your capstone is that performance.

Prompt refactoring means rewriting a messy prompt into a clear structure — same meaning, better clarity.

• Break the prompt into blocks: role, context, constraints, output format.
• Remove the noise.

Think of a messy instruction:

• Messy: "go outside, bring something, also do work, come fast".
• Clean: 1. go to shop, 2. buy milk, 3. take change, 4. come home.

Same meaning — the structure makes all the difference.

A prompt testing checklist is a set of questions to verify prompt quality before you use it.

• Like inspecting a bike before buying — check the engine, brakes, and fuel.
• A prompt must be checked before delivery.

A good prompt is written, a great prompt is tested. Apply the checklist to every prompt before you use it.

The prompt improvement loop is: Ask → Check → Refine → Retest. Your first prompt is never perfect.

• Every iteration improves the output.
• The same way you adjust milk and sugar when making tea.

Prompt debugging is finding the mistake and fixing the prompt properly using the process Diagnose → Explain → Rewrite → Test. Broken prompts are the best learning tool — real skill comes from practice.

The mechanic example: the customer says "bike not working"; the mechanic checks, finds the issue, fixes it, and explains the solution. That is exactly how a prompt engineer handles a failing prompt.

A meta-prompt is a prompt that works on another prompt. The Prompt Debugger is an AI that analyzes a prompt, finds issues, and rewrites an improved version automatically.

• With a meta-prompt you are not just using AI — you are building AI tools.
• Normal prompt: solves the task. Meta-prompt: improves the prompt itself.

Real-world simulation means applying all debugging skills to a practical scenario:

• A client says their chatbot is giving wrong answers.
• As the prompt engineer, you analyze, identify, refactor, test, and present the solution.

Each step has a clear job and a clear output:

• Step 1 → produces the explanation.
• Step 2 → uses it to produce a summary.
• Step 3 → uses the summary to produce a quiz.

Clean and controllable.

Decomposition means breaking a big task into small, clear steps in the correct order. Use the framework Input → Process → Output:

• Each step receives an input.
• Does one process.
• Produces one useful output.

The market example:

• "Go bring groceries" → confusion.
• Making a list, checking the budget, buying the items, and verifying → a clear process.

Break the task, define each step's output, and connect the outputs.

Format locking means keeping the same structure across all steps:

• If Step 1 gives output as headings and bullets, Step 2 must receive and produce the same structure.
• Consistency between steps is what keeps a pipeline stable.

Notice each step tells AI exactly what format to produce. That is format locking — it stops the pipeline from drifting into a different structure halfway.

A branching pipeline changes its path based on a condition. The logic is simple: IF condition → Route A, ELSE → Route B.

• Different users get different outputs.
• AI becomes smart and personalized.

First, AI classifies the user; then it sends them down the correct path.

• Marks ≥ 80 → advanced class.
• Marks < 80 → basic revision.

Define the condition clearly and keep the two flows separate.

Controlled chaining means maintaining consistency and control across every step — same topic, same format, same rules, no unwanted content.

• Without it, AI can drift away from the topic.
• It can add new information you never asked for.

• Same topic, same format, same rules in every step.
• Each step uses only the previous output.

The story example:

• Start a story and suddenly change the topic → confusing.
• Staying on the same story → the message stays clear.

In every step, use only the previous output and repeat the rule "do not add new information".

Variable injection means using placeholders in a prompt so the same prompt works for many inputs. Instead of writing a new prompt every time, you build one template with variables like {{TOPIC}}, {{USER_INPUT}}, and {{JOB_ROLE}} — then pass real values.

• Identify the dynamic parts.
• Replace them with variables.
• Reuse the template for any input.

The birthday-template example:

• "Happy Birthday {{NAME}}" — change the name, keep the same template.
• That is variable injection: identify the dynamic parts, replace them with variables, and reuse the template.

A reusable workflow template is a pre-built prompt system you can reuse for many tasks. It is built from:

• Small modules — each step does one job.
• Variables — {{TOPIC}} style placeholders.
• Standard structure — the same output shape every time.

Build once, use many times.

The school-template example:

• Every student writes name, roll number, and subject on the same form — same format → reusable.
• Your prompt template works the same way — change the topic, keep the system.

JSON is the format used to pass structured data between workflow steps. Instead of messy text, each step returns a clean JSON object, and the next step reads it.

• JSON keeps the data organized.
• Tools and APIs can read it automatically.

The capstone is the final real-world simulation that uses everything you learned — chaining, pipelines, variables, structure, and automation thinking. It is your journey from learner to system designer.

• Pick one real problem.
• Design the workflow, break it into steps, and define inputs and outputs.
• Apply chaining and variables, then test with multiple inputs.

• Pick one real problem.
• Design the workflow, break it into steps, and define inputs and outputs.
• Apply chaining and variables, then test with multiple inputs.

Optimization means improving a workflow for better performance — speed, accuracy, cost (tokens), and stability. The first workflow is never perfect:

• Review it.
• Remove unnecessary steps.
• Reduce token usage.
• Finalize the template.

The travel example:

• First time → the long route.
• Next time → the shorter one.

Same destination, better efficiency. Optimize your workflows the same way.

Prompt serialization means structuring prompts in an organized format before sending them to AI — instead of one plain sentence, you use structured roles:

• System — behaviour.
• User — task.
• Assistant — response.

Structured prompts give control; plain prompts are simple but weak.

The system message controls behaviour (how AI should act); the user message gives the task (what AI should do). System = how, User = what.

• And remember: the system message has higher priority.

The school example:

• A friendly teacher gives a soft explanation.
• A strict teacher gives a direct one.

Same question, different answers — because the behaviour changed. Change the system message and the whole output changes.

The Alpaca template is a structured format used to train AI models, with three clear parts:

• Instruction — what to do.
• Input — additional context.
• Output — expected result.

It makes datasets consistent and improves instruction understanding.

The homework example:

• Instruction = write about planets.
• Input = focus on Earth and Mars.
• Output = 5 points.

Clear structure gives a clear answer — whether for a student or an AI model being fine-tuned.

ChatML is the structured format for conversations in systems like ChatGPT. It uses roles — system, user, assistant — so AI can remember context across multiple turns.

• Enables real multi-step interaction.
• Best for chatbots and conversations.

The INST format wraps the instruction inside special tags and is used by LLaMA and open-source models. It is a lightweight format with no roles — just the instruction inside [INST] ... [/INST].

The exam-sheet example:

• A question placed inside a box is easy to identify.
• The instruction inside tags works the same way — the model knows exactly where the task begins and ends.

Prompt minimalism means writing short, clear, and precise prompts without losing meaning. It is not just short — it is short + clear + complete.

• Long prompts are not always better.
• Extra words create confusion.

The shop example:

• Confusing: "go to the shop near the temple, check the milk, if fresh buy, maybe biscuits also…".
• Clear: "Buy one fresh milk packet from the nearby shop".

Cut the extras, keep the meaning.

Prompt noise is extra words that do not add value to the task. Noise removal means cutting those words while keeping the meaning the same.

• AI does not need polite fillers like "please kindly".
• Fillers only add tokens and confusion.

A token is a small piece of text that AI reads — AI processes your prompt token by token, roughly word by word.

• More tokens → more processing → more cost.
• Short prompts = low tokens = low cost.

The bus-ticket example:

• Short distance → low cost.
• Long distance → high cost.

In AI: short prompt = low tokens = low cost. In real projects every API call costs money — minimize tokens without losing meaning.

Consistency means getting the same style and format every time.

• Always control the output format so the result looks identical across runs.
• Without it, AI gives random, unpredictable output.

Safety means reliable and cautious responses. A hallucination is AI giving an incorrect but confident answer.

The directions example:

• Three people give three different answers.
• So you ask: "Are you sure? What is your assumption? Can you verify?"

Ask AI the same questions.

Key rules for reliability:

• Control the output format.
• Ask AI to show its assumptions.
• Ask for sources when needed.
• Ask for verification.
• Allow AI to say "Not sure".
• Use self-consistency checks.

Multimodal AI understands and creates multiple types of data — text, image, video, and audio.

• It can look at a photo and answer questions about it.
• It can turn a text description into an image or a video.

AI is no longer text-only.

Image prompting means describing a scene in text to generate an image. More details give a better image.

• Formula: Subject + Environment + Lighting + Camera + Style.
• Each element adds quality to the final image.

The boy example:

• "A boy studying" → a flat image.
• Add the village house, warm evening sunlight, camera angle, and realistic style → a professional-looking scene.

Cinematic prompting controls the visual feel using lighting + camera + mood:

• Lighting — controls the atmosphere.
• Camera — controls the perspective.
• Mood — controls the emotion.

Small changes create a big impact.

The same person photographed in morning sunlight looks happy; in dark shadows they look serious.

• Same subject, different feeling — that is the power of lighting, camera, and mood.

Video prompting means describing a moving scene to generate a video. The formula adds action and camera movement to the image formula:

• Formula: Subject + Action + Camera Movement + Lighting + Style.
• A photo is a standing person; a video is that person walking.

Negative prompting means telling AI what NOT to generate. A normal prompt says what you want; a negative prompt says what you don't want.

• Essential for image and video generation.
• AI may add unwanted things like blur, extra fingers, or bad lighting — the negative prompt removes them.

These words solve the most common AI image problems — reuse them in every image or video prompt:

Example for a realistic image:

Always use negative prompts for: AI images, videos, realistic faces, and product photography.

PAL means Program-Aided Language prompting. Instead of asking AI to answer directly, you ask AI to write a program and solve the problem. The thinking becomes Language → Code → Answer.

• Most people ask AI like a guesser.
• Smart people ask AI like an engineer.

The basic formula is extremely simple — "Write a Python program to + [your problem]". For bigger problems, add the details and ask for step-by-step output.

PAL is powerful because the code is accurate, step-by-step, and reusable.

• Helps: calculations, finance, real estate, data analysis, business planning, automation.
• Not needed: stories, motivation, marketing content — places where no logic is required.

Real-estate profit:

EMI calculation:

Monthly revenue (total + average):

Find the topper — works for 1000 students too:

Investment growth: