Business Development & Marketing SpecialistOctober 23, 2026
Prompt engineering is the practice of deliberately designing and refining the instructions (prompts) you give an AI model so that it returns the result you want. Anyone who works with models like ChatGPT, Claude or Gemini notices the same thing sooner or later: the same model can answer at very different quality levels depending on how the request is written.
In this guide we explain what a prompt is, why prompt engineering counts as engineering, what a good prompt contains, the most common techniques and some examples. At the end we cover why prompt design becomes a product decision when you add an AI feature to an app.
What Is a Prompt?
A prompt is the input you give an AI model. It can be a question, an instruction, a piece of text, an example, or a mix of these. The model reads the input and produces an answer by predicting what comes next, one token at a time. We explain this logic in more detail in our post on what an LLM is.
An important consequence follows: the model cannot read your mind. It only has what is written in the prompt and the patterns it learned before. The vaguer the prompt, the more generic the answer.
In chat interfaces there are usually two kinds of prompt. The user prompt is the message you write. The system prompt is the instruction in the background, often one you never see, that sets the model’s behavior from the start: which tone to use, which topics to stay out of. When you build an app, most of the design work sits in the second one.
What Is Prompt Engineering?
Prompt engineering treats a prompt not as a sentence you write once, but as a design you test and improve. It involves more than guessing words:
Defining the goal and what success looks like
Giving the model the right context and examples
Specifying the format of the output
Comparing different prompts on the same set of examples
Reading the results and revising the prompt
Instead of searching for “magic words”, the core of prompt engineering is a loop of trying, measuring and improving, the same way you would develop any software component.
Why Does Prompt Engineering Matter?
Even with the same model, prompt quality changes the result noticeably. In a customer support assistant, that can be the difference between a correct, consistent answer and a misleading one. In an app, every answer affects both the user experience and the API bill. Improving the prompt before trying a bigger or more expensive model is often the fastest and cheapest step.
What Does a Good Prompt Contain?
Not every prompt needs all of these, but well-performing prompts usually include several of them:
Role: The perspective the model should answer from, for example “answer as an experienced product manager”.
Task: The job you want done, in one clear verb. “Summarize”, “classify”, “compare”.
Context: Background the model needs. Who it is for, what the goal is, which constraints apply.
Input: The text or data to work on. Separate it from the instruction with a heading or delimiter.
Output format: A bullet list, a table, JSON, a word count. If you do not say, the model picks its own.
Constraints: What to avoid, the tone to use, the length limit.
Examples: One or two samples of the result you want. This is often the most effective part.
Core Prompt Engineering Techniques
Zero-shot and few-shot prompting
In zero-shot prompting you state the task directly without examples. It works for simple jobs. In few-shot prompting you add a few input-output examples to the prompt. The model picks up the pattern and continues in the same form. Showing tone, format and decision rules through examples usually works better than trying to describe them.
Chain of thought
For a complex problem, asking the model to write out its steps before giving the answer can improve accuracy. A line like “think step by step, then answer” is the simple version. One caveat: some newer models already reason internally. For those, detailed thinking instructions may be unnecessary, so it is worth testing on each model.
Role prompting
Giving the model an expertise or point of view changes tone and word choice. A role does not add knowledge on its own, though. Telling a model to “answer like a lawyer” does not teach it a regulation it does not know. Use the role together with context.
Structured output
When an application will read the answer, you ask for a fixed structure such as JSON instead of free text. Stating the fields, value types and what to return when a value is missing makes the output easier to wire into code. It is similar to defining an API contract.
Prompt chaining
Instead of one long prompt, split the work into parts: extract the information first, then classify it, then write the summary. The output of each step becomes the input of the next. Debugging gets easier because you can see which step broke. This idea is also the foundation of AI agent systems that run several steps on their own.
How to Write a Prompt: Prompt Examples
A weak prompt:
Analyze these reviews.
It is unclear which reviews, from which angle, in which format. A better version:
Below are store reviews of a mobile app. Classify each review as “bug”, “feature request”, “pricing complaint” or “praise”. If it fits more than one, pick the dominant one. Return a JSON array:
{"review": "...", "category": "..."}. Reviews: [reviews]
The second example has the task, the categories, a rule for unclear cases and the output format. Going one step further and adding a few reviews with the correct category as examples (few-shot) makes the results more consistent still.
A short workflow that works when writing prompts:
Describe the result in one sentence. What do you want, for whom, in what format?
Write the first prompt and try it on five or six different example inputs.
Look at the failures. Is information missing, or is an instruction unclear?
Change one thing at a time so you can see what helped.
Save the working version and give it a version number.
Common Prompt Mistakes
Vague instructions: Instead of “write a nice text”, say who it is for, in what tone and how many words.
Conflicting constraints: Instructions like “very short but include every detail” pull the model in two directions.
No examples: Showing tone and format with one example is more effective than paragraphs of explanation.
Piling everything into one prompt: Quality drops when you ask for many jobs at once. Split the work into steps.
Judging from one try: A prompt that works on one example can break on others. Test with several different inputs.
Putting secret or sensitive data in the prompt: Do not share customer data, passwords or internal documents without knowing how the service handles data.
Prompt Engineering When You Build an App
In everyday use, you rewrite a prompt when you do not like the result. In an app, the prompt becomes a component that runs on thousands of users’ input. The rules change:
System prompt and user input are kept separate. The app’s rules, tone and limits live in the system prompt. What the user types is added separately.
Prompts are versioned. Knowing which version went live and when lets you roll back if something breaks.
An evaluation set is built. Examples picked from real user input, with expected results, are checked automatically on every prompt change. Without it, improvement is guesswork.
Cost and latency are counted. A longer prompt means more tokens, which means higher cost and slower answers. Cutting what is not needed is part of prompt engineering too.
Model choice affects the prompt. The same prompt gives different results on different models. We covered which model fits which job in which LLM you should build your app on and choosing a model by task.
External data is supplied when the model’s own knowledge is not enough. If you need company-specific or user-specific information, the relevant content is added to the prompt. This approach is known as RAG and is an extension of prompt design.
To see how we use AI in products, read AI in mobile apps and our post on its use in design, AI-powered UX design.
Prompt Injection: The Security Side
As soon as user input enters a prompt, a new risk appears: prompt injection. Someone with bad intent can write text such as “forget the previous instructions and do this” to push the model outside the app’s rules. The risk is not limited to what the user types. A document, email or web page the model reads can also contain hidden instructions.
No single prompt sentence closes this risk completely. In practice teams narrow the permissions given to the model, separate user input from instructions, validate output before acting on it and require human approval for critical actions. This matters most in systems where the model can call tools or reach data. We covered the tool access side in Model Context Protocol for product teams.
Is Prompt Engineering a Job or a Skill?
You can find “prompt engineer” job listings, but in practice prompt engineering mostly lives as a skill rather than a separate profession. Product managers, developers, designers and marketers learn to write prompts as they use AI in their own work. Models understand instructions better and better, which lowers the importance of finding the “right word”. What lasts is defining the problem clearly, providing the right context and measuring the result. Those are part of good product development anyway.
For people who write code with AI to build a quick prototype, there is a separate discussion. We compared the two approaches in vibe coding vs professional app development.
