Most disappointing AI answers are not caused by a bad model. They are caused by a short question that leaves the AI guessing. An AI chatbot only knows what is in the conversation: it does not know who you are, what you already tried, who the answer is for, or what you plan to do with it. Fill in those gaps and the same assistant suddenly seems much smarter.
You do not need special tricks, secret phrases or a "prompt engineering" course. You need the same habits you would use when asking a busy, knowledgeable colleague for help. This guide covers those habits, with concrete before-and-after examples you can adapt.
When you type "write a cover letter", the AI has to invent everything you left out: the job, your experience, the tone, the length. It fills those blanks with the most typical choices, which is why the result reads like a template.
A good prompt removes the guesswork. It usually answers four questions:
You will not need all four every time. "What is the capital of Canada?" is a perfectly good prompt. But for anything you plan to use, such as an email, a plan, an explanation or a decision, the extra twenty seconds pays off.
Vague verbs produce vague answers. "Help me with" or "tell me about" can mean almost anything. Say exactly what you want done: write, summarise, compare, explain, list, rewrite, check.
Weak: "Tell me about budgeting."
Strong: "I earn about $4,000 a month after tax and keep running out of money before payday. Suggest a simple budgeting method I can set up in one evening, and explain how to track it without an app."
Context is anything that would make a human expert give you a different answer. Your level of knowledge, your location, your budget, your deadline, and what you have already tried all count.
Weak: "How do I get rid of mould?"
Strong: "There is black mould on the ceiling of my bathroom, about the size of a dinner plate. I rent, the room has no window, and I want to clean it safely myself. What should I use, what should I avoid, and when should I tell my landlord instead?"
Mentioning that you rent and that there is no window changes the advice completely: ventilation becomes the real issue, and the landlord question is worth raising.
"Explain how interest rates affect inflation" will get you a textbook paragraph. "Explain it to a 14-year-old using a pocket-money example" or "explain it for someone who just started a finance degree" will each get you something you can actually use. Naming the audience is one of the quickest ways to improve an explanation.
If you do not ask for a format, you get the AI's default, which is often several paragraphs with a summary on the end. Ask for what you will actually use: a table, a checklist, three options, a numbered set of steps, or a draft under a certain word count.
Weak: "Compare electric and petrol cars."
Strong: "Compare a small electric car and a small petrol car for someone who drives about 40 km a day, mostly in the city, and can charge at home. Use a table with purchase cost, running cost, maintenance and range, then give a two-sentence recommendation."
Length limits are especially useful for writing. "Under 120 words" or "three short paragraphs" stops the AI from padding, and it is easier to ask for more than to cut a long draft down.
If you want help with an email, a paragraph, a recipe or a set of notes, paste the actual text. "Make my email sound more professional" without the email forces the AI to write a new one from scratch. With the email in front of it, it can keep your facts and fix only the tone.
Weak: "Can you make my complaint email better?"
Strong: "Here is an email I want to send to an airline about a lost bag. Keep all the facts and dates, make it firm but polite, and end with a clear request for what I want them to do. [paste email]"
Before pasting, remove anything sensitive the AI does not need, such as booking references, account numbers or other people's contact details. Placeholders like [NAME] or [BOOKING REF] work just as well.
AI models can state wrong things in a confident, fluent voice. For anything factual you plan to rely on, such as prices, rules, dates, health information or anything recent, ask the AI to show where its information comes from, and then open those sources.
Weak: "What are the rules for carrying medication on a plane?"
Strong: "I'm flying from Australia to Japan next month with prescription medication. Search for the current rules for bringing it into Japan, list what documents I might need, and link to the official sources so I can check them."
If an answer includes no sources, treat specific numbers and rules as a starting point to verify, not a final answer.
The first answer is a draft. Chatbots remember the conversation, so you can steer instead of rewriting your whole prompt. Short follow-ups work well:
Telling the AI why something did not work is more useful than just saying "try again". "Too generic, I'm a beginner who has never cooked fish" gives it something to fix.
Long conversations have a limit to how much the AI can keep in view, and many assistants only send the most recent part of a long chat with each new message. If you change topic, or a conversation has gone on for a long time and the answers start ignoring earlier details, start a fresh chat and restate the important points.
| Weak prompt | Stronger prompt |
|---|---|
| "Dinner ideas" | "Five weeknight dinners for two adults, under 30 minutes each, no pork, using mostly supermarket basics. List the ingredients for each." |
| "Fix my CV" | "Here is my CV summary for a junior data analyst role. Rewrite it in three sentences, keep it factual, and highlight my Excel and SQL experience. [paste]" |
| "Plan a trip to Italy" | "Plan 7 days in Italy in May for two people who like food and walking but not museums. We land in Rome and fly out of Milan. Day-by-day, with train travel between cities." |
| "Explain this error" | "I'm learning Python. Here is my code and the full error message. Explain what the error means in plain language and show the smallest change that fixes it. [paste]" |
If you want a starting point, this structure works for most everyday requests:
"I need [task] for [audience or purpose]. Context: [the facts that matter, constraints, what you've tried]. Please give me [format] in about [length]. [Optional: Include sources / list any assumptions you make.]"
Asking the AI to "list any assumptions" is a useful habit: it shows you where it filled in gaps, so you can correct them in your next message.
The fastest way to learn is to take one question you asked an AI recently, rewrite it using the steps above, and compare the answers. You can try these prompts free in Tobava AI, which searches the web and shows its sources when a question needs current information. For more on features like follow-ups, personalisation and temporary chats, see the Tobava AI chat guide.