Each spin unlocks one expert-crafted prompt matched to the AI you land on. 50 prompts across ChatGPT, Claude, Gemini and Perplexity — productivity, coding, research, writing, career and more.

The Three Layers of AI

Artificial Intelligence (AI) is any computer system that performs tasks requiring human-like intelligence — recognizing speech, translating languages, making decisions, creating art. It is the broadest category.

Machine Learning (ML) is a specific approach within AI. Instead of programming exact rules ("if email contains 'win a prize', mark as spam"), you show the computer thousands of examples and it learns the rules itself. This is how Netflix recommends shows and Gmail filters spam.

Deep Learning (DL) is a subset of ML that uses neural networks with many layers. Each layer detects increasingly complex patterns — from edges to shapes to objects to meaning. ChatGPT, image recognition, and self-driving cars all use deep learning.

Analogy: Think of AI as "vehicles." ML is "cars" (one type of vehicle). DL is "electric cars" (one type of car). All electric cars are cars, all cars are vehicles — but not all vehicles are electric cars.

How Neural Networks Learn

A neural network is loosely inspired by the human brain. It has layers of connected nodes (neurons). When you train it:

Step 1: Show it a picture of a cat labeled "cat." It makes a random prediction (probably wrong).

Step 2: Calculate how wrong it was (the "loss").

Step 3: Adjust millions of tiny numerical weights to be slightly less wrong next time.

Step 4: Repeat with millions of pictures. Each repetition makes tiny improvements. After billions of adjustments, it becomes excellent at recognizing cats — and dogs, cars, faces, cancer cells, etc.

Key insight: The network was never told "cats have whiskers and pointy ears." It discovered these patterns itself from the data. This is the magic of machine learning.

Supervised vs Unsupervised vs Reinforcement Learning

Supervised learning: You provide labeled examples. "This email = spam. This email = not spam." The model learns to predict labels for new data. Used in: medical diagnosis, price prediction, spam filtering.

Unsupervised learning: You provide data without labels. The model discovers hidden patterns. "These customers behave similarly — they might be a distinct segment." Used in: customer segmentation, anomaly detection, recommendation systems.

Reinforcement learning: The model learns by trial and error, receiving rewards for correct actions. This is how AlphaGo learned to beat the world champion at Go — by playing millions of games against itself. Used in: game AI, robotics, self-driving cars.

What is a Parameter?

Parameters are the learnable numbers inside an AI model. Think of them as tiny dials — each one controls how the model processes information. GPT-1 had 117 million parameters. GPT-3 had 175 billion. GPT-4 has roughly 1.8 trillion. More parameters generally means more capability, but also requires more computing power and energy to run.

What is Training Data?

Training data is the information used to teach the model. For ChatGPT, this included a significant portion of the internet — books, articles, websites, code repositories, and more. The quality and diversity of training data directly determines the quality of the AI. Biased training data produces biased AI. This is one of the biggest challenges in the field.

What is a Transformer?

The Transformer is the architecture (blueprint) behind all modern AI language models. Introduced by Google in 2017 in a paper called "Attention Is All You Need," it revolutionized AI by solving a fundamental problem: how to process text efficiently.

Before Transformers: AI read text one word at a time, left to right (like a human reading slowly). This was painfully slow and the model would "forget" the beginning of long texts by the time it reached the end.

After Transformers: AI processes all words simultaneously, with each word "looking at" every other word to understand context. This is massively parallel and dramatically faster.

The Attention Mechanism — The Key Innovation

Consider the sentence: "The cat sat on the mat because it was tired." What does "it" refer to? The cat or the mat? Humans know instantly — "it" means the cat. But how does AI know?

Attention solves this. Each word calculates a relevance score with every other word. "It" gives high attention to "cat" and low attention to "mat." This lets the model understand relationships, references, and meaning across the entire text.

Self-attention formula: For each word, compute three things: a Query (what am I looking for?), a Key (what do I contain?), and a Value (what information do I carry?). Compare each Query with all Keys. The highest-scoring Keys get the most attention. This is computed for every word simultaneously.

How ChatGPT Actually Generates Text

ChatGPT generates text one token at a time. Given all the tokens so far, it predicts the probability of every possible next token, then picks one (influenced by the temperature setting). Then it adds that token to the input and predicts the next one. And so on, token by token, until it reaches a stopping point.

Critical insight: ChatGPT does not "think" or "understand." It predicts statistically likely continuations. It has no internal model of truth or reality. This is why it can write a convincing essay about a topic it is completely wrong about — the text sounds right even when the facts are wrong. This is hallucination.

What is a Token?

AI models do not read words — they read tokens. A token is a piece of text: sometimes a whole word, sometimes part of a word, sometimes a character. "Hello" = 1 token. "ChatGPT" = about 2 tokens. "Artificial Intelligence" = about 2-3 tokens.

Why tokens matter: API pricing is per-token ($5/M for GPT-5.5 input). Context window limits are in tokens (1M for GPT-5.5). Non-English languages (Hindi, Arabic, Chinese) typically require MORE tokens per word, making them more expensive.

Quick math: 1,000 words ≈ 1,333 tokens. A typical email (200 words) ≈ 267 tokens.

Context Window

The context window is the total amount of text (in tokens) that a model can "see" at once. It includes everything: system prompt, your messages, the AI's responses, and any documents you paste.

In 2023, context windows were 4K-32K tokens. In 2026, frontier models have 1M+ token windows — enough to process entire codebases or books. Llama 4 Scout has a 10M token window.

Limitation: Research shows models can struggle with information in the middle of very long contexts (the "lost in the middle" problem). Information at the beginning and end gets more attention than information buried in the middle.

Temperature

Temperature controls randomness. At temperature 0, the model always picks the single most likely next token — giving deterministic, focused output. At temperature 1.5, it considers less likely tokens too, producing creative, varied, sometimes surprising output.

Use low (0.1-0.3) for: factual questions, code, data extraction, anything where accuracy matters.

Use high (0.8-1.2) for: creative writing, brainstorming, poetry, marketing copy.

Default (0.7) is good for most tasks.

What is Hallucination?

Hallucination is when AI generates plausible-sounding but factually incorrect information with full confidence. It invents studies that do not exist, cites cases that were never filed, and generates quotes nobody ever said. This happens because the model predicts likely text, not truthful text.

How to protect yourself: Always verify critical facts. Use Perplexity for research (it cites sources). Ask the AI "are you confident about this?" Use RAG (retrieval-augmented generation) to ground AI in real data.

What is RAG?

Retrieval-Augmented Generation is a technique where the AI searches a knowledge base before generating an answer. Instead of relying only on training data (which can be outdated or wrong), it retrieves relevant real documents and uses them as context. This dramatically reduces hallucination and enables current, accurate answers about your specific data.

The Prompt Structure Framework

Every great prompt has up to 6 parts. You do not need all 6 every time, but knowing them lets you craft the right prompt for any situation:

1. Role: Who should the AI pretend to be? "Act as a senior data scientist specializing in time-series forecasting with 10 years at Fortune 500 companies."

2. Context: What background does the AI need? "I run a small e-commerce store selling handmade candles. My monthly revenue is ₹2 lakh. I have 3 employees."

3. Task: What specifically should it do? "Analyze my last 6 months of sales data and identify the 3 most actionable trends."

4. Format: How should the output look? "Present as a numbered list with each trend as a heading, 2-3 sentences of explanation, and one recommended action."

5. Constraints: What should it avoid? "Do not use technical jargon. Keep total response under 400 words. Do not suggest hiring more staff."

6. Examples: What does good output look like? Show 1-2 examples of the format and quality you expect.

Before and After: Real Prompt Improvements

The weak prompt gives the AI no context, no format, no constraints. It will produce generic advice. The strong prompt gives a role, context, specific task, format, constraints, and a follow-up instruction. The output will be dramatically more useful.

The Iteration Loop — How Professionals Prompt

Professional prompt engineers never expect the first output to be perfect. They use a feedback loop:

Round 1: Send your initial prompt. Read the output carefully.

Round 2: "Good start, but make the tone more casual and add a specific example for point #3."

Round 3: "The example is too technical. Replace it with something a small business owner would relate to."

Round 4: "Perfect. Now condense the whole thing to half the length without losing any key points."

Each round refines the output. This iterative approach produces dramatically better results than trying to craft one perfect prompt. Think of it as sculpting — you rough out the shape first, then refine the details.

Common Mistakes That Kill Prompt Quality

Mistake 1: Being vague. "Write something about marketing" → The AI has to guess everything. Be specific about topic, audience, format, length, and tone.

Mistake 2: Asking for too much at once. "Write a complete business plan with financial projections, marketing strategy, and operational plan" → Break this into 4-5 separate prompts, each building on the previous output.

Mistake 3: Not providing examples. If you want a specific format or style, showing one example is worth more than 100 words of description.

Mistake 4: Ignoring negative constraints. Telling AI what NOT to do ("do not use bullet points", "avoid jargon", "no more than 200 words") is often more powerful than positive instructions.

Mistake 5: Not iterating. Accepting the first output as final. Always refine.

Tokens = money and speed. Every word in your prompt costs tokens. Here's how to get more for less:

Choose a PDF file, then press Convert. All languages (Hindi, Arabic, Chinese, Tamil, etc.) are preserved. Your file never leaves your device.

Choose a .docx file, then press Convert. Hindi, Arabic, Chinese, Tamil and every other language render using your browser's real fonts — nothing is approximated or skipped.

In the print dialog set Destination to “Save as PDF”, then Save. Real Hindi / Arabic / Chinese / Tamil glyphs are kept because your browser draws them, not a built-in PDF font.