Why does ChatGPT (or Gemini, or Claude) always give me generic answers?

Almost always because the prompt is too generic. If you ask something vaguely, the AI has to guess what you mean and picks the most 'average' answer possible. Add context, a precise goal, and the output format you want.

Do I still need to write 'think step by step' in my prompts?

With the most recent models, which already do internal reasoning before answering you, this trick often isn't needed anymore and can sometimes make the answer worse instead of better. It still makes sense mainly with simpler or 'fast' models, not with flagship ones.

Is it worth writing extremely long prompts with every possible detail?

No. An overly long, disorganized prompt is hard to follow even for an AI. A prompt structured into clear sections (context, request, format) works better than one long, stream-of-consciousness block of text.

How to write effective prompts for artificial intelligence

You know that moment when you ask ChatGPT something and the answer is correct but… flat? Generic, as if it were written for anyone except you? The problem is almost never the tool — it’s the prompt: the text you write to ask the AI for something. In this guide I’ll explain how to write better prompts, with practical examples you can copy and adapt right away.

What a prompt actually is

A prompt is simply the instruction you give an AI chatbot: a question, a request, a task to carry out. The quality of the answer depends heavily on the quality of that instruction, for a simple reason: the AI can’t read your mind, it only works with the information you give it in the prompt (and, if it exists, the conversation before it).

If you write “write me a social media post”, the AI has to guess almost everything: about what, for which platform, in what tone, how long. The result will inevitably be generic. If instead you write “write me an Instagram post announcing the opening of an artisan gelato shop in Bologna, friendly and informal tone, maximum 80 words, with two or three emoji”, you get something you can use almost immediately.

The 4 components of a good prompt

There’s no magic formula, but a good prompt tends to have four elements, even if you don’t always need all four together:

1. Context. Who you are, what situation you’re in, why you need this. “I’m writing my economics thesis” is a different context from “I’m preparing a social media post”, even if the following request were identical.

2. A precise task. What you want to get, specifically. Not “tell me about marketing” but “list 5 low-cost marketing strategies for a small brick-and-mortar shop”.

3. Output format. How you want the answer structured: a bulleted list, a table, continuous text, a certain word or paragraph count. If you don’t specify it, the AI chooses for you, and it’s often not the format you actually needed.

4. Examples (when they help). If you have a specific style in mind, show an example of it. “Write like this example: [paste an example]” works far better than a thousand adjectives trying to describe a style in words.

A practical trick: structure the prompt into sections

For more complex requests, instead of writing everything in one paragraph, try splitting the prompt into labeled sections. The model follows instructions more reliably when they’re organized this way, rather than scattered across one block of text:

CONTEXT: I run a small e-commerce store selling pet products.
TASK: Write the product description for a felt cat bed, medium size.
TONE: Friendly but professional, without overselling.
FORMAT: One paragraph of 60-80 words, followed by 3 bullet points with the main features.

You don’t need to follow this template to the letter every time — for a simple question it would be overkill — but for more involved tasks (an article, an analysis, a technical document) it saves you quite a few attempts.

”Think step by step”: is it still needed?

Until recently, one of the most common tricks was adding phrases like “think step by step” or “explain your reasoning before answering” at the end of a prompt, to push the model to be more accurate on complex problems. With the most recent, more advanced models, which already run an internal reasoning phase before giving you an answer, this trick is often no longer necessary — and in some cases you can get worse results, because you’re duplicating work the model is already doing on its own.

The practical rule today is: give the task and the constraints clearly, and let the model decide how to reason through it. If you’re instead using a “fast” or lighter model (the kind built for instant answers, not deep reasoning), explicitly asking it to go step by step can still help.

Iterate, don’t expect a perfect answer on the first try

A common mistake is treating a prompt like a command the AI must answer perfectly on the first attempt, a bit like a search engine. It works better if you treat it as the start of a conversation: make a request, look at what you get, then correct it with a follow-up message (“too long, cut it in half”, “use a less formal tone”, “add a practical example to the second point”). It’s almost always faster and more precise than rewriting the original prompt from scratch trying to anticipate every detail in advance.

Common mistakes to avoid

Being too vague. “Help me with my resume” gives much worse results than “rewrite this resume bullet to be more concise, highlighting measurable results instead of generic responsibilities”.
Not specifying the audience or purpose. The same topic needs to be explained differently to a beginner versus an expert: say so explicitly in the prompt.
Endless, disorganized prompts. More information doesn’t automatically mean a better prompt: if the text turns into an unstructured stream of consciousness, the model struggles to tell what’s actually a priority.
Not reviewing the answer critically. Even the most carefully crafted prompt doesn’t protect you from errors or inaccuracies in the answer: always verify important data, figures and claims before using them.
Pasting sensitive data into the prompt. Never paste personal data, confidential work information, or private documents into an AI chat, unless you know exactly how that data is handled by the service you’re using.

A before-and-after example

Weak prompt: “Write me an email for an unhappy customer.”

Effective prompt: “Write a reply email to a customer complaining about a delivery that arrived 5 days later than promised. Professional but empathetic tone, acknowledge the mistake without over-apologizing, offer a 10% discount on their next order as an apology. Maximum 120 words, close with a question asking whether the package still arrived in good condition.”

The difference in the result you get is huge, and the time spent writing the second prompt is far less than the time you save by not having to rewrite a generic answer from scratch.