What prompt engineering is, and why bother if the model answers anyway
"I tried it and I'm no good at it" is the most frequent sentence from people who have already opened a model. The cause is not the phrasing but two things: how the task was set and which model was chosen. Below are the six parts of a proper brief and the three techniques that deliver most of the value.
"I tried it and I'm no good at it" is the most frequent sentence from people who have already opened a model. The cause is not the phrasing but two things: how the task was set and which model was chosen. Below are the six parts of a proper brief and the three techniques that deliver most of the value.
---
"I tried it and I'm no good at it"
The most frequent sentence from people who have already used AI. They opened it, asked, got a general answer, shrugged.
Two conclusions follow and both are wrong. Either "the model is overrated" or "I must be doing something wrong, I should learn". The second is nearer the truth but does not say what to learn.
The difference between people it works for and people it does not lies not in secret phrasings. It lies in two things: how the task was set and which model was chosen.
What a prompt actually is
A prompt is not "a query to a neural network". It is a brief for somebody who does not know the context.
Imagine you hired a person who can do anything and has not the faintest idea of your situation, your client or your taste. What would you say to them? Clearly not "write me some copy".
That is prompt engineering: learning to set a task so the person does not have to guess.
What a properly set task consists of:
The role. Who the model should be. A sales specialist and a proofreader give different answers to the same question.
The context. What the situation is, who it is for, what has already happened. Most bad answers come from missing context rather than bad phrasing.
The task. What specifically has to be done. One action rather than five at once.
The format. What has to come out: a list, a table, text, how long. Without it you will get whatever the model finds convenient.
Constraints. What not to do, which words not to use, what to avoid. An underrated part — prohibitions work harder than instructions.
The material. What to work from. A prompt with your text attached works fundamentally more reliably than one where the model retrieves everything from memory.
The main mistake: one model for every task
Most people take one model — usually the best known — and try to do everything in it.
That is like hammering nails with a screwdriver because it came to hand first. Sometimes it works.
Models have different characters. One holds long context better and works with structure. Another is stronger at conversational text. A third searches faster and pulls in current data. A fourth handles code better.
The difference between "it works" and "it doesn't" is often not in the prompt at all — it is that the person opened the wrong model.
Hence a second skill more important than phrasing: knowing the landscape and understanding what to open for what.
The three techniques that give most
Of everything called prompt engineering, three things deliver most of the value.
1. Supply the material rather than asking it to remember
The worst mode: "tell me about X". The model retrieves from memory and can be wrong confidently and invisibly.
The best: "here is a document, go through it against these points". When all the material has been supplied by you, there is nothing to invent — and reliability rises several times over.
2. Prohibit rather than only request
"Write some copy" gives mediocre copy. "Write some copy, do not use the words 'unique', 'innovative', 'in today's world', do not open with general reflections, do not sum up at the end" gives something else entirely.
A list of prohibitions is almost always more useful than a list of wishes.
3. Break the task into steps
One big prompt saying "do everything" loses to a chain where each step's result feeds the next. First the analysis, then the structure, then the text, then the clean-up.
That is the same principle ready assistants and skills are built on: they are not one prompt but a sequence.
An example: one task, two requests
The difference is clearer than any explanation.
The usual way: "Write a post about our services."
What comes out: smooth text for nobody, with turns of phrase like "in today's world", a list of advantages and an invitation to get in touch. People scroll past that.
The right way:
> Role: a copywriter writing for owners of small offline businesses.
> Context: the audience is car repair shop owners, half of them have no website and do not think they need one. Their real pain is that customers cannot find prices and ring with the same question.
> Task: one post showing that pain and offering a solution.
> Format: up to 900 characters, short paragraphs, no list of advantages.
> Constraints: do not use the words "unique", "innovative", "in today's world". Do not open with general reflections. Do not promise sales growth.
> Material: [attach your own examples of customer questions].
The second request is several times longer, but it gets written once and is reused from then on: only the material block changes. That is the difference between asking and setting a task.
What not to learn
Memorising "magic phrases". Lists of magical formulations from the internet go out of date with the models and work worse than a clearly described task.
Chasing length. A long prompt does not mean a good one. In some tools — search ones particularly — a short query works better than a detailed one, because you are not getting in the tool's way.
Taking the result on trust. A separate skill more important than prompting: verify everything verifiable against sources. A model names non-existent things as confidently as real ones, and in text they are indistinguishable.
When a prompt is not needed at all
An important part people skip.
Prompt engineering solves the task of "explaining to the model what to do". But it need not be solved if the task has already been described by somebody.
A ready assistant is a task already set. It has a role, it has its own questions for you, it has an output format. You describe the situation in ordinary words and it asks for what is missing itself.
A skill is a chain of tasks. Not one prompt but a sequence where each step's result goes onwards: the copywriter writes, the designer presents, the proofreader checks.
Hence the practical conclusion: first check whether a ready one exists. A prompt gets written when the task is non-standard. For a standard one it is faster to take what is already assembled — and the result is usually better, because whoever assembled it has done it a hundred times.
Where to start
The "Prompt engineering" express course — a foundational course on working with all the main models. Not a list of techniques but two things: a map of the landscape — which model for what — and a universal language for setting a task that works in any of them.
The result: you know what to open for a specific task and can set it so you are not redoing the answer five times.
The price is $9. The course is included in the Basic plan.
Registration is free and opens three days of full Basic access — meaning you can simply take this course having paid nothing. Alongside it are five more foundational courses, the ten-lesson sales module and chats with 140+ AI assistants across 14 categories.
About the assistants separately, because they change the question. A ready assistant is a task already set. You open a chat on your subject — marketing, sales, HR, legal, analytics — and it asks what it needs itself. No prompt has to be written at all.
Prompt engineering is needed where no ready solution exists. For everything else it is faster to take a ready one.
Do one thing today: take the last poor answer you got and rewrite the request to the scheme — role, context, task, format, constraints, material. The difference is usually visible on the first attempt, and it explains everything else.
---
Sixteen methods in four groups
Prompt engineering is not a bag of tricks but a systematised set of techniques. Module 4 holds sixteen of them across four groups.
Fundamentals and context — how to set the role, the frame and the material so the model does not build the missing parts itself.
Logic and reasoning — how to make it show its working rather than only its conclusion. Critical where a mistake has to be visible in the process.
Multi-step processes — how to break down a large task so every step solves its own sub-question.
Autonomous methods — where the model chooses what to do next itself.
And module 3 covers separately the three things people start with: tone of voice, role prompts and cascading prompts.
The phrasing from the cascades lesson: you lead the model along a clear plan, like a scriptwriter.
Modules 3 and 4 — the Basic plan.
---
What to read next
[Six brief fields](/en/blog/six-brief-fields) — what is missing from a specific request.
[How to choose an AI tool for a task](/en/blog/how-to-choose-ai-tool) — the other half of the cause.
[How Claude differs from other models](/en/blog/claude-vs-other-models) — when structure matters more than phrasing.
[How to check what a model tells you](/en/blog/how-to-verify-ai-output) — a skill more important than prompting.
[Best open LLMs 2026](/en/blog/open-source-llm-guide) — the full guide to open models.