Prompt tricks that do not work

A video tells you there is a secret sentence. Type it before your question, and the AI will answer like a professor. You try it. The answer comes back longer, and you cannot tell if it is any better.

Most of those tricks have now been tested properly. This lesson tells you what the tests found. It also tells you where to put the effort you save.

The tricks you keep hearing about

Six of these come up again and again. You have probably seen at least three.

• Telling it that it is a world-class expert.
• Offering it money as a tip for a good answer.
• Saying that your job depends on the answer.
• Threatening it with something bad if it gets the answer wrong.
• Being extremely polite, with please and thank you in every line.
• Adding think step by step to the end of every question.

Researchers have run these through proper tests. Not one question, but hundreds of hard exam questions, repeated many times, on several AI systems.

Take the tip and the threat first. One study used a set of 198 hard science questions and a set of engineering questions. It tried offering money. It tried threats. The finding was that threatening or tipping a model has no significant effect.

The expert role went the same way. A second study from the same group tested six AI systems. It gave each one a matching role, such as you are a physicist. Then it asked hard physics questions. There was no steady gain. Only one system out of six moved at all.

That study tried weak roles too, like you are a child. Those made the answers worse.

One more thing is worth knowing. On any single question these tricks can swing the score a long way, up or down. Nobody can tell in advance which way it will go. Across a full set of questions the swings cancel out. That is luck, not a technique.

A role changes the voice, not the facts

Now the fair part, because it is easy to be unfair here. Giving the model a role does change something real. It changes the words it picks and the tone it uses.

You are a patient school teacher.
Explain simple interest to a class 8 student.
Use short sentences and one example from a village shop.

Look at which lines are doing the work. The last two carry all of it. They say who the reader is, what the topic is, and what the answer should look like.

The first line only sets the voice. That is genuinely useful when you want a gentler tone or plainer words. It does not make a date, a rule or a sum more correct. Hold on to that difference and no prompt video will fool you again. It sits on top of how to ask so you get a useful answer, which is the real skill.

Please and thank you cost nothing

So should you be rude instead? No.

The evidence here is genuinely mixed, and you deserve to hear that. One 2025 test asked the same 50 questions in five different tones. Very polite prompts scored 80.8 percent. Very rude prompts scored 84.8 percent. Rude came out slightly ahead.

A larger 2026 study pointed the other way. It looked at 22,500 prompts and replies across five AI systems. The languages were English, Hindi and Spanish. Polite wording improved English answers by up to 11 percent. In Hindi, indirect and respectful phrasing did best of the tones they tried. That is one study, not a rule.

No single tone wins everywhere. So write the way you would speak to a person. Please costs you nothing and it does no harm. Being rude buys nothing you can count on, and it is a poor habit to build.

Think step by step belongs to older models

This one was real once. That is exactly why it spread so far.

A few years ago, most models answered straight away, in one go. Telling them to think step by step made them work through the problem first. On some of those older models the gain was more than ten percent on hard questions.

Then the models changed. Many now do that stepping by themselves, before they show you anything. On these newer thinking models the measured gain fell to about three percent. On one system it was slightly negative.

The cost is not small. Those extra words made answers take 20 to 80 percent longer on thinking models. On a budget phone with a small data pack, that is your time and your money for almost nothing.

There is a better move. When a question needs many steps, switch on the thinking or reasoning setting in the app. What thinking mode really does explains when the extra wait is worth it.

Keep show me the working for maths homework, when you want to see the steps yourself. That is a different job, and a good one.

One test for any tip you see

New tips will keep arriving. You need a test you can run yourself, in one second.

Your class, your district, your exam, your budget, your crop. All of that is information. The model could not have known any of it. Your own notes are information. Saying what a good answer looks like is information.

Telling it that it is world-class is not information. Neither is a tip, a threat, or a third please. Those add praise or pressure. There is nothing in them the model can actually use.

Information in, a better answer out. Flattery in, the same answer with more decoration on it.

Spend the effort here instead

Here is where those same five minutes pay you back.

Give your situation. Your class, your district, your exam, your budget in rupees, your phone. In a study of college students in a prompting contest, this was one of the few things with a clear measured gain. It is also the easiest part for you, because you already know all of it.

Give it your own material. A photo of the page, your notes, the text from an official website. Put that material above your question, not below it. Giving it your own book, notes or photo is the biggest single jump in answer quality.

Say what a good answer looks like. One line, before you ask. If you cannot describe a good answer in a sentence, the model cannot produce one.

Ask it to admit uncertainty. The model was trained on tests where a guess scores better than a blank. So it guesses, in a calm and confident voice. Give it permission not to know.

If you are not sure about something, write NOT SURE.
Do not guess.

Ask for a shape. A table with named columns. Five steps. Under 200 words. Asking for the answer in the shape you need shows you how to do that well.

Put together, a strong request looks like this. There is nothing magic anywhere in it.

I am in class 10 in a village in Uttar Pradesh.
I want to understand simple interest for my board exam.
Explain it in under 200 words. Keep sentences under 15 words.
Then give me 3 practice sums, with the answers at the end.
If you are not sure about something, write NOT SURE.

A good answer to that comes back short, in easy words, with the sums at the bottom. If it comes back long and hard, say so and ask again. Asking again is normal, not a failure.

Nobody needs to sell you a prompt list

You will see prompt packs for sale. A thousand prompts for your exam. Five hundred prompts to earn money. Many are copied from other lists that were written for older models.

Copying is not even the biggest problem. Long copied prompts go stale. The companies that build these models now advise short, fresh prompts for each new model. They also warn that piling on extra instructions can make behaviour worse. So a list written for last year’s model can drag today’s answers down. Reading an old tutorial safely goes further into this.

One survey of this field lists 58 named prompting techniques for text alone. Almost none of them are meant for you. Five habits cover nearly everything an ordinary person needs. Context, your own material, a clear picture of a good answer, permission to say NOT SURE, and a shape.

Why do the flattery tricks fail at all? The answer is in what the machine is really doing when it replies. It is not deciding to try harder for a polite person. The next lesson, it is guessing the next word, shows you that whole mechanism. After that, most of these tricks stop even sounding believable.