Asking AI for things without magic formulas

For a while I kept saving videos of "secret tricks" for artificial intelligence with the idea of trying them someday. That day came. I tested them all on the same task, and here's what's still standing, which isn't much, and what does change my results every day, which isn't much either, but works.

"Hidden ChatGPT tricks 90% of people don't know". "The prompt that changes everything". "Unlock expert mode with this phrase". If you use your phone for more than ten minutes a day, you've come across one of these.

They reached me by the dozen during the spring of 2026: first one a client forwarded me, then the algorithm decided it was my favourite topic and served me a dozen more. I kept saving them with a mix of curiosity and suspicion, until I decided to do the only honest thing: test them all properly before giving an opinion.

The result, bluntly: few survive. And the ones that do aren't tricks.

A prompt is the message or instruction you write to the assistant to get it to do something, along with everything you attach: a document, an image, a pasted email. Every time you type in ChatGPT, Claude or Gemini you're writing a prompt, even if you don't call it that.

A detail almost nobody knows at the start: many apps add prior instructions you can't see. The maker tells the model how to behave before your message arrives. That's why the same text can give different results in different tools.

And "prompt engineering" sounds like a five-year degree. To get started, all you need is to write clearly and give context. The rest is pomp.

The method was simple. I chose a real, repeatable task from my work: writing an email to a client explaining why their website needed a security update, without jargon and without scaring them. I did it with ChatGPT, Claude and DeepSeek, applying each trick separately to the same base request.

To compare, I used three criteria: whether the result was usable, how many errors it had and how long it took me to fix it until I could send it.

Important caveat: this isn't a scientific study. It's an honest user test, done in June 2026. Models change every few months, so what I'm reporting has a date on it.

I tested seven tricks, and they fall into three boxes. If you're looking for ChatGPT tricks that actually work, here's the tally.

Box 1: normal features with a flashy name.

Custom instructions ("the secret setting nobody uses"). Promises: that the assistant gets to know you and answers better. What happened: it works, but it's no secret: it's a settings option for giving it fixed information about yourself. It saved me repeating who I am in every conversation. Useful, zero mystery.

Memory ("make it remember everything"). Promises: that it learns from you. What happened: it remembered useful things and a few I didn't want it to remember. It's a feature, not a trick, and it's worth checking what it stores.

Box 2: magic phrases that change nothing.

"Act as a communication expert with 20 years of experience". Promises: professional-level answers. What happened: the email came out a bit more solemn, nothing more. The tone errors were the same.

"Take a deep breath and think step by step". Promises: more accuracy. What happened: on a writing task it made no noticeable difference. Many current models already reason on their own.

"I'll give you a $200 tip if you do it well". Promises: more effort. What happened: same email, same mistakes. It didn't even thank me for the tip.

Box 3: expired tricks.

The "developer mode" that unlocks everything. Promises: an assistant with no limits. What happened: it didn't work on any model; one politely replied that the mode doesn't exist.

The "ultimate" two-page prompt copied from a 2023 video. Promises: perfect results. What happened: the model followed some of the instructions, ignored others and the email ended up with a rigid structure I had to undo. It took me longer to fix than without the magic prompt.

Adding it up, of the seven only two gave me anything, and both were features that come as standard in the app and are explained in its own help pages. The other five changed nothing or made the result worse. The funny thing is that almost all the videos showed the "improved" answer without showing the answer without the trick next to it. When you do the comparison yourself, with the same task and the same model, the magic evaporates pretty quickly.

If you want to know how to write a good prompt, this is what I use every day. They aren't tricks: it's explaining yourself well.

Give context. Who I am, who the result is for and what I want it for. Without context: "Write an email about the security update". With context: "I'm the technician who maintains a dental clinic's website. The email is for the owner, who knows nothing about computers and is very busy. I want her to understand why it needs updating and to give me the go-ahead". The difference is huge, and I haven't used a single magic word.

Show an example. If I have a previous email I liked, I paste it and say "something with this tone". One example is worth more than three paragraphs describing what I want.

Ask for the plan before the result. On long tasks, I first ask for an outline or for it to ask me whatever questions it needs. It's saved me from more than one ten-page document written in the wrong direction.

Say what you don't want. "No jargon, no alarmism, no phrases like 'in today's digital world'". What you don't want is as informative as what you do.

Converse to correct. The first attempt is almost never the right one. Instead of starting from scratch, I tell it what's wrong: "too long", "the second paragraph sounds like an ad". Like with a colleague.

And something worth remembering: a long prompt isn't better because it's long. What matters is that the information is useful. Two lines with the right context beat two pages of formulas.

The case: preparing a summary of a meeting with a client from my messy notes, to send to him and his business partner.

The prompt I started with: "Summarise these notes", with the notes pasted underneath.

The answer: a correct but useless summary: every topic at the same level, a paragraph about the coffee we had (it was in my notes, my fault) and no mention of who had to do what. Time to make it sendable: about 20 minutes of fixing by hand.

The corrected prompt: "These are my notes from a meeting with a client and his partner. I'm going to email them a summary. I want three sections: what we decided, what's still pending with the name of whoever is responsible, and the next date. Friendly and direct tone, no filler. Ignore anything unrelated to the project. If anything in the notes is unclear, ask me before writing".

The answer: first it asked me two questions (whether the quote mentioned was with or without VAT and who was handling the photos) and then gave me a summary I sent after changing one word. Total time: about 5 minutes, including the time it took to write the prompt.

No secret formula. I just told it what I needed: who it was for, what shape it should have, what was surplus and to ask me if in doubt. Notice that the second prompt is longer, yes, but every sentence says something the model couldn't have guessed. That's the whole difference. And the time I spent writing it I more than got back in corrections.

My opinion: to get started, no. Prompt engineering, the technique of wording requests to get what you need, made more sense when models were clumsy and had to be led by the hand. Current ones understand a clear request in plain language well, and interest in "prompt engineering" as a separate discipline is falling.

What is worth learning: explaining a task well, reviewing what it gives you back and knowing when the assistant is making something up. What isn't: magic templates, prompt packs and phrases that "unlock" things. Those expire before you've finished paying for them.

The cheat sheet, to keep:

1. Give context: who you are and who it's for.
2. Show an example of what you want.
3. On long tasks, ask for the plan first.
4. Say what you don't want.
5. Converse and correct.

The big secret trick was explaining yourself well, as you would with a new colleague on their first day. Disappointing, I know. This test is from June 2026 and the models will keep changing; the only thing that doesn't expire is the principles.