Talking to AI So It Actually Helps (Prompting for Normal People)

Part 4 of the series "AI, Without the Hype". Start at Part 1.

You get far better results from AI by doing four simple things: tell it who it is, give it context, show it one example of what good looks like, and state the format you want back. Add "think step by step" for anything with logic, and that is essentially all of practical prompting. No course, no secret words, no jargon. Most people get weak answers because they type a vague half-sentence and expect the machine to read their mind. It can't.

I build AI products in Ahmedabad and prompt these models all day. The difference between a useless answer and a genuinely good one is almost never the model — it is the instruction. Here is the version a busy, non-technical person can actually use.

The direct answer

In Part 2 we saw that AI predicts the most likely next words based on what you give it. So the logic is simple: better input narrows what counts as a likely answer. Vague in, vague out. Specific in, useful out.

You are not "programming" the AI. You are briefing it — exactly like a sharp new assistant on their first day. You would not just say "write something." You would say who it's for, what you need, and what good looks like. Do that with AI and the quality jumps immediately. Everything below is just five concrete ways to give a good brief.

The 5 moves

Remember them as role, context, example, format, step-by-step. You rarely need all five at once. Even two will transform most answers.

• Role — tell it who it is. Start with "You are a..." This pulls the model toward the right kind of language and knowledge. "You are a CA who explains tax to small shop owners" produces a very different answer from a blank request.
• Context — give it the boring details. It cannot see your situation. Who are you, who is this for, what's the goal, what matters, what to avoid. The "obvious" facts in your head are invisible to the AI until you type them.
• Example — show one sample of "good." This is the most underused move and the most powerful. Paste one example of the style, tone, or structure you want, and the model copies the pattern. One good example beats three paragraphs of description.
• Format — say exactly what you want back. A table? Five bullet points? A 50-word reply? An email under 100 words? If you don't say, you get the model's default, which is usually too long and too generic. Name the shape.
• Step-by-step — for anything with logic. For maths, planning, comparisons, or decisions, add "think step by step before you answer." It nudges the model to lay out its reasoning instead of blurting a guess — and you can see where it went wrong.

Everyday before/after examples

Theory is cheap. Here is the difference these moves make for ordinary people.

A shopkeeper writing to a supplier. Before: "Write a message to my supplier about late delivery." After: "You are helping a kirana store owner in Ahmedabad. Write a polite but firm WhatsApp message in simple Hindi to a supplier whose rice delivery is 4 days late. Keep it under 60 words. Ask for a new delivery date today." The first gives a generic English paragraph. The second gives something he can send as-is — right language, right length, right tone.

A student studying for an exam. Before: "Explain photosynthesis." After: "You are a patient teacher. Explain photosynthesis to a Class 8 student in 5 short bullet points, using a simple everyday example. No technical words without explaining them." The first gives a textbook wall. The second gives something a child absorbs.

A small-business owner writing a customer email. Before: "Write an email apologising for a delay." After: "You are the owner of a small custom-furniture business. Write a warm, honest email to a customer whose dining table is delayed by one week. Apologise, give the new date (next Friday), and offer free delivery as a sorry. Under 100 words. Here's the tone I like: [paste one past email]." The first is robotic. The second sounds like a real human who runs a real shop — because you gave it a role, the facts, a format, and an example.

Notice the pattern across all three: you are not learning tricks. You are simply telling the AI what you already know in your head.

What is NOT worth obsessing over

There is a lot of prompt folklore online. Most of it is noise. Save your energy.

• "Magic words" and secret phrases. Stacking flattery like "you are the world's greatest expert" does little. Today's models do not need to be buttered up. A clear brief beats incantations.
• Perfect grammar and spelling. The model understands messy, casual, typo-ridden input fine. Write like you talk. Clarity matters; polish does not.
• Absurdly long prompts. More words is not better. Relevant words are. A tight, specific four-line prompt usually beats a rambling page.
• Copying giant templates you don't understand. A template you can't adjust is a cage. Learn the five moves and you can write a good prompt for anything, instead of hunting for someone else's.

If you want to go deeper on the craft once the basics are second nature, I keep a fuller prompt engineering guide. But honestly, the five moves cover the vast majority of real-world needs.

Where prompting hits its limit

Prompting is powerful, but it is not magic, and it is worth knowing the wall you will eventually hit.

No prompt can give the AI information it was never trained on. You cannot prompt your way to last night's match score, your company's private files, or this morning's price list. The knowledge simply isn't in the model, and as we saw in Part 3, asking anyway just invites a confident guess.

When you genuinely need the AI to work with specific, private, or current information, prompting alone is the wrong tool. Two other approaches take over:

• RAG (retrieval). You give the model the right documents to read at answer time — your files, your data, the current facts — so it responds from real material instead of memory. Getting the right information in front of the model is its own discipline, which I cover in context engineering.
• Fine-tuning. You further train a model on your own examples so it deeply absorbs a specific style or task. Heavier, slower, and only worth it for narrow, repeated jobs.

For 95% of what a normal person does day to day, you will never need either. The five moves are enough to turn AI from a frustrating toy into a genuinely useful assistant.

If you remember one line from Part 4, make it this: don't make the AI guess what you want — brief it like a smart new assistant, and it will act like one.

Next up — Part 5: where your data actually goes when you use AI.