Prompt Engineering 101

The quality of AI output depends heavily on how you ask. A vague prompt — “write me an email” — gets a vague answer that could be anything. A specific, well-structured prompt — “write a 120-word apology email to a customer whose order was delayed two days, offer a 15% discount, warm and professional tone” — gets something you can almost paste into Gmail and send. This skill is called prompt engineering, and it’s the single most practical AI skill you can learn today.

The idea sounds too simple: tell the AI clearly what you want. But anyone who has watched a coworker type “summarize this” into ChatGPT and get a useless paragraph back knows it’s harder than it looks. A good prompt is more like a mini-brief than a wish. You specify who the output is for, what tone it should use, how long it should be, what to include, what to leave out. In this mission, you’ll learn to write prompts that get you there on the first try — or the second, when the first falls short.

The phrase “prompt engineering” barely existed before 2020. Older AI — the kind running Google Translate or the spam filter inside Gmail — didn’t take prompts at all. You gave it input, it gave you a fixed output, end of story. There was nothing to “engineer.”

That changed when large language models (LLMs) arrived. Around 2020, models like GPT-3 showed that a single model could write poems, answer questions, summarize articles, and write code — as long as you asked in the right way. Researchers noticed something strange: two people asking the same question in different words would get very different quality answers. Adding “let’s think step by step” to a math question made models markedly more accurate. Saying “You are an expert lawyer” before a legal question changed the whole tone and depth of the reply.

A small group of researchers started cataloging these tricks. Their papers had names like “chain-of-thought prompting” and “few-shot learning.” By 2023, when ChatGPT reached 100 million users faster than any consumer product in history, prompt engineering moved from research labs to daily practice. Companies hired “prompt engineers.” Schools argued about whether to teach it at all.

The real shift wasn’t technical — it was cultural. For the first time, ordinary people could talk to software in natural language and get something useful back. But “natural language” turned out to be harder than it sounded. We leave huge amounts of context unsaid when we ask a friend for help. AI has no such context. Every detail you leave out, it fills in with its best guess — and that guess defaults to generic, because generic is the safest statistical average of its training data. Prompt engineering is the habit of putting the context into the question, so the model doesn’t have to guess.

A strong prompt almost always has five moving parts, even if you don’t name them. Once you can spot them, you can fix a weak prompt by adding the missing ones.

Role — Who should the AI be? “You are a customer support agent.” “Act as a strict code reviewer.” One line shifts vocabulary, tone, and what the model pays attention to. It doesn’t give the model new knowledge, but it narrows the range of plausible answers. A financial question asked to “a cautious CFO” sounds very different from the same question asked to “a startup founder” — even from the same model.

Task — What exactly do you want? “Write,” “summarize,” “rewrite,” “compare,” “extract.” Use a specific verb. “Help me with this” is not a task — it’s a vibe. Strong prompts lead with a clear verb and object: summarize this article, rewrite this paragraph in plain English, extract the three action items from this meeting note.

Context — What does the AI need to know that it can’t guess? Your audience (“my parents, who aren’t technical”), your situation (“the client is already annoyed”), or your goal (“to get them to reply by Friday”). Context is what separates a generic answer from one that fits your actual life. Most prompts fail here — people describe the task but forget the context.

Format — How should the answer look? “In three bullet points.” “As a table with columns for pros and cons.” “JSON with keys: name, price, url.” The model will happily produce a wall of text if you don’t specify. Format is also how you make the output scannable and reusable, especially when you’ll copy it into another tool.

Constraints — What’s off-limits or required? Word count (“under 100 words”), language (“in Thai”), vocabulary (“no jargon”), structure (“start with a question”). Constraints prevent the most common AI failure — the answer that is technically correct but longer, more formal, or more padded than you needed.

Beyond the five parts, three classic techniques are worth knowing. Few-shot prompting means giving the AI 2–3 example inputs and outputs so it learns the pattern from examples rather than description alone. Chain-of-thought is adding “think step by step” or “show your reasoning” to complex problems — it often lifts accuracy noticeably on math, logic, and analysis. Iterating is the habit of refining one part at a time: if the tone is off, tweak the role line; if the length is wrong, tighten the constraint. Don’t regenerate the whole prompt when you can surgically patch it.

You don’t need all five parts every time. A quick brainstorm can skip format. A short translation can skip role. But when an AI output lands flat, the debug question is simple: which part did I leave out?

Prompt engineering is the foundation. Once you can reliably get AI to produce what you actually want, you’re ready to use it as a creative partner rather than a toy. The same skill that helps you write a better email also helps you brief an image generator, debug code with ChatGPT, summarize a long research paper, or draft a Thai-language Instagram caption that doesn’t sound machine-translated.

But there’s a catch: a well-prompted AI can still produce something convincing that is wrong, off-brand, or inappropriate for your context. Prompting gives you quality input; it doesn’t guarantee quality output. That’s why even the best prompters still read every result, fact-check the claims, and adjust the tone before sending.

In the next mission, “Create with AI Tools”, we’ll zoom out from text to the full AI toolkit: text models like ChatGPT and Claude, image models like Midjourney, design tools like Canva AI, translators like DeepL, and code helpers like GitHub Copilot. Picking the right tool for the job matters as much as writing a good prompt — using a text model to generate an image, or an image model to write a headline, wastes time no matter how clean your prompt is. The skill you just built — asking clearly, adding context, iterating — carries across all of them.