How to Write AI Prompts That Actually Work (With Examples)

Most "prompt engineering" guides tell you the same four things: be specific, assign a role, give context, state the format. All true. All useless on their own, because nobody shows you what that actually looks like when you sit down to type.

So this guide does the opposite. You get a fill-in-the-blank template you can paste into any AI tool today, four complete before/after prompt pairs, and — the part almost nobody teaches — what to do when the first answer comes back wrong. That last skill matters more than the perfect opening prompt.

Why Your Prompts Give Generic Answers

AI models are not mind readers. They respond to exactly what you write, and when you leave something unspecified, the model fills the gap with the most average, safest possible choice. That is where "AI-sounding" output comes from — not from the model being weak, but from you not making a decision the model then had to make for you.

Every part you specify is one less thing being guessed. That is the whole theory. Now the practice.

The 5-Part Prompt Template

Copy this and fill in the brackets. It works in ChatGPT, Gemini, Claude, or any other text AI:

ROLE:    You are a [specific role with relevant experience].
TASK:    [One clear verb] [exactly what you want].
CONTEXT: This is for [audience]. Background: [what the AI can't guess].
FORMAT:  Give me [structure: a table / 5 bullets / 300 words / numbered steps].
RULES:   [Constraints: tone, what to avoid, reading level, length limits.]

You do not need all five every time. But each one you skip is a decision you are handing to the model. Here is what each part is really doing:

Part	What it fixes	Weak vs strong
Role	Depth and vocabulary level	"Explain X" → "You are a teacher explaining to a 15-year-old"
Task	The AI doing the wrong job	"Help with this" → "Critique this and list what's unclear"
Context	Generic, off-target answers	Nothing → "For beginners in Nepal with no paid tools"
Format	Walls of text you must re-read	Nothing → "A table with 3 columns"
Rules	Wrong tone, wrong length	Nothing → "Under 200 words, plain English, no jargon"

The most underrated one is Task. "Draft", "critique", "summarize", "compare", and "analyze" are five completely different jobs. Choosing the right verb changes the answer more than any other single word in your prompt.

Four Before and After Examples

1. Getting feedback on your writing

• Before: "Can you improve this paragraph?"
• After: "You are an editor for a beginner-focused tech blog. Review the paragraph below and identify where it is unclear, repetitive, or unsupported. Do not rewrite it — list the problems as bullets, then suggest one stronger structure. Audience: readers who are not native English speakers."

Why it works: "improve" invites a generic rewrite. Naming the job (identify problems, don't rewrite) gets you something you can actually learn from.

2. Learning a hard topic

• Before: "Explain APIs."
• After: "You are a patient teacher. Explain what an API is to someone who has never coded, using one everyday analogy. Then give one real example of an API a blogger might use. Keep it under 200 words, plain English, no jargon unless you define it."

3. Planning something

• Before: "Give me blog post ideas about AI."
• After: "Give me 10 blog post ideas about free AI tools for students. For each, give the title and one line on the search intent it answers. Titles must be under 60 characters and phrased as questions people actually type. Avoid topics that require paid tools."

4. Working with your own data

• Before: "Summarize this."
• After: "Summarize the text below into three sections: key decisions, action items, and open questions. Use bullets. If something is ambiguous, list it under open questions rather than guessing."

That last rule — "list it rather than guessing" — is a quiet trick worth reusing everywhere. It gives the model permission to admit uncertainty instead of inventing something confident and wrong.

The Real Skill: Fixing a Bad Answer

Here is what separates people who get good at AI from people who give up: they treat the first answer as a draft, not a verdict. When the reply is wrong, do not start a new chat and re-type everything. Correct one thing at a time:

• Too generic? "That's too general. Rewrite it for [specific audience] and include one concrete example."
• Too long? "Cut this to 150 words. Keep the examples, remove the introduction."
• Wrong tone? "Rewrite in plain, direct English. No marketing language."
• Missing the point? "You focused on X. I need Y. Try again focusing only on Y."
• Not sure it's right? "What assumptions did you make? What might be wrong here?"

Change one variable per attempt. If you change three things at once and the answer improves, you have learned nothing about which change worked. This is the same discipline as image prompting, which I covered in my guide to AI image prompts — one motion per clip, one change per iteration.

Three Mistakes to Stop Making

1. Asking a question when you want a task. "What do you think about my headline?" gets you an opinion. "Give me 5 stronger alternatives and explain why each is better" gets you work.
2. Stuffing everything into one prompt. Ten requirements in one paragraph means the model will drop half. Break big jobs into steps and build on each answer.
3. Accepting facts without checking. Confident wrong answers are the number one AI trap. For anything factual, ask for sources or verify independently — especially for numbers, dates, and names. If the AI is working from your own documents, tools like the NotebookLM workflow I use for studying cut this risk sharply because answers come with citations.

Practical Takeaway

Save the 5-part template somewhere you can paste it from — a notes app, a pinned message, anywhere. For the next week, use it for one real task per day and keep the prompts that worked. Within days you will have a personal prompt library, which is worth far more than any paid prompt pack you can buy. Good prompting is not a talent or a secret. It is just the habit of making your decisions explicit instead of hoping the model guesses right.

FAQ

How do I write a good AI prompt as a beginner?

Use five parts: role, task, context, format, and rules. State who the AI should be, exactly what job you want done, background it cannot guess, the structure you want back, and any limits on tone or length. Then refine one thing at a time.

Why does AI give me generic or useless answers?

Because anything you leave unspecified gets filled in with the most average option. Vague prompts produce vague output. Adding audience, format, and constraints usually fixes it immediately.

Do the same prompts work in ChatGPT, Gemini, and Claude?

Mostly yes — the five-part structure transfers across tools. You may need small adjustments to tone or detail since each model has different default habits, but a clear prompt is clear everywhere.

Should I use "act as a..." role prompts?

Yes, when the role genuinely changes the answer — a teacher, an editor, a reviewer for beginners. Skip it for simple factual lookups where it adds nothing. Roles shape depth and vocabulary, not accuracy.

How do I stop AI from making things up?

Ask it to flag uncertainty ("list anything ambiguous instead of guessing"), request sources for factual claims, and verify numbers and names yourself. For work based on your own documents, use a tool that answers only from your sources with citations.