A prompt is the instruction you give an AI model such as ChatGPT or Claude to get a particular output: a question, a task or a detailed set of working instructions. Prompt engineering, or prompting, is the skill of writing, testing and adjusting those instructions until the output is usable.
How precise your prompt is determines how much you still have to correct afterwards. For professionals, prompting is therefore not a trick for one-off questions but part of a way of working in which AI output is checked before it goes any further. If you want to learn to work with AI in a structured way, the AI courses offer an approach for each tool.
At the bottom of this page you will find all published articles in the prompt engineering cluster, one of the sections of the AI knowledge base.
What is a prompt?
A prompt is the text instruction you give an AI model to generate a specific output. That instruction can be a question, a task, a description or a structured instruction of several hundred words.
The word prompt has three meanings. As an adjective, prompt means punctual or immediate, as in “a prompt payment”. In computing, the prompt is the symbol on the command line that shows the computer is waiting for input. In the context of AI, a prompt is the input a language model responds to, and prompt and AI prompt mean the same thing.
At its core, a prompt is nothing more than asking AI a good question, only with more structure and direction. The difference between a prompt and prompting is like the difference between a sentence and the skill of writing. A prompt is the building block; prompting is the practice of putting those blocks together and refining them.
A prompt can be as short as a four-word question (“summarise this document”) or as long as a structured document with a goal, context, format and constraints. In everyday AI work you see both extremes and everything in between. Which form suits which task depends on what you want to achieve and how predictable the output needs to be.
What does prompt engineering mean?
Prompt engineering means writing instructions for an AI model and refining them until the output fits what you need. The term is used for large language models (LLMs) such as ChatGPT, Claude, Gemini and Grok.
Each model responds slightly differently to the same prompt. Prompting is therefore not a uniform technique but a skill you adapt to the tool you work with.
Some people reserve the term prompt engineering for the more technical side of prompting, particularly API use and automated applications, and call the everyday version simply prompting. For most professionals that distinction does not matter. Both terms refer to the same basic idea: stating clearly what you want, in a form the AI model can work with.
In a production context, with API work and automated pipelines, prompting does become more technical, because reproducibility and cost play a bigger role than in one-off conversations.
Since ChatGPT became widely available at the end of 2022, prompting has developed along two lines. The models have become more capable, so they produce good output from less detailed prompts. At the same time, new types of model have appeared, notably reasoning models, which call for a different approach.
The net effect: prompting has not become simpler, but it has matured. What was sold as a collection of tricks in 2023 is now a structured skill that professionals actively train.
I moved from one-off prompts to workflows myself after incorrect output ended up in production. Output that looks right at first glance but turns out to be wrong when you check it forces you to treat prompting as part of a way of working, not as a trick.
How prompt engineering works in practice
Prompting works by giving an AI model context and direction. Based on patterns in its training, an LLM predicts which text logically follows the input. The more specific and detailed the input, the more accurately the model can make that prediction.
A vague prompt produces a wide range of possible outputs; a precise prompt narrows that down to what you need.
Many users do not expect this: the same prompt can give different output at different times. LLMs are not deterministic. There is an element of chance in the way they generate text, which means variation can occur even with identical input.
For one-off questions that is not a problem; for reproducible workflows it is. This behaviour follows from the way AI models work in general; how they generate text and why output varies is covered in AI basics.
A second mechanism behind prompting is that small changes in the prompt can have a large effect on the output. Adding one sentence of context, or naming a specific audience, can shift an answer from generic to usable.
That makes prompting iterative: you write, look at what comes out, adjust and repeat until the output fits what you need.
The five building blocks of a good prompt
The five building blocks of a good prompt are goal, context, format, constraints and quality check. This structure makes you decide what you need before you write the instruction itself. In practice professionals rarely use all five at once, but each building block helps with specific kinds of task.
The goal describes what you want to achieve with the output. Not “write a text”, but “write an invitation to a team meeting about the Q3 results”. The context gives the AI model the background it needs to make the output relevant: who is it for, what prompted it, what knowledge can the model assume.
The format sets out what the output looks like: an email, a list, a table, a short paragraph, a detailed report. Constraints are limits or requirements you state explicitly, such as a maximum word count, a formal tone, or facts only and no opinions.
The quality check is the least used but most important building block. How do you know whether the output is right? This is the step many professionals skip, and it is what separates AI as an experimental toy from AI as a professional working tool. Letting output go out unchecked is exactly how errors end up in publications.
The difference between a vague prompt and a well-built one is clearest side by side, for the same task:
Vague prompt: “Write an email to a client about the delay.”
Well-built prompt: “Write an email to a long-standing client of our accountancy firm, with the aim of informing the client and keeping their trust (goal). The 2025 annual accounts will be two weeks later than agreed, because the client did not supply the bank details until 12 September (context). No more than 150 words in three short paragraphs, with the new delivery date in the last paragraph (format). No apologies for a delay that is not our fault and no legal commitments (constraints). Finally, check that every date in the email matches the details above (quality check).”
The first prompt produces a generic email that you still have to fill in yourself. The second produces an email you only need to check for facts and tone. The quality check in the prompt does not replace your own check: a model sometimes copies a date incorrectly and then reports that everything is correct.
Which prompt engineering techniques are there?
Six prompt engineering techniques deliver the most in everyday work: zero-shot prompting, few-shot prompting, chain-of-thought prompting, role prompting, prompt chaining and giving context to an LLM. For each technique, the sections below say when it makes sense.
Zero-shot prompting
Zero-shot prompting is the simplest form: you give an instruction without examples and ask the model to answer straight away. For many everyday tasks this works fine, especially when the task falls well within the general capabilities of an LLM.
“Summarise this text in five bullet points” is a typical zero-shot prompt. Zero-shot works well for standard tasks; for more specific tasks you are more likely to get generic output.
Few-shot prompting
Few-shot prompting adds a few examples to your prompt to show the model what kind of output you expect. Instead of only describing what you want, you show it.
For classification, extracting structured data and writing in a fixed house style, few-shot prompting gives better output than a description alone, because format and style are hard to capture in words. Two to four examples are usually enough; more examples often add little and make the prompt unnecessarily long.
Chain-of-thought prompting
Chain-of-thought prompting is the technique of having the model reason step by step before it gives an answer. For more complex tasks, especially analyses and multi-step problems, chain-of-thought prompting gives noticeably better results.
With reasoning models, chain of thought is less necessary, because those models already work through problems step by step internally. According to Anthropic’s model documentation, thinking is always on for Claude Opus 5.5. In ChatGPT you choose a reasoning level in the model picker (Medium or High, and on Pro also Extra High). The technique remains valuable for classic models and for situations where you explicitly want to see how the model reaches a conclusion.
Role prompting
Role prompting means giving the AI model a specific area of expertise by assigning a role in your prompt, such as senior SEO specialist, lawyer, copy editor or tax adviser. Giving a role shifts the tone, the level of detail and the kind of considerations in the output.
Role prompting works well when you want output that matches a specific expertise, and less well when you use the role as a gimmick with no real link to the task; “pretend you are a Nobel Prize winner” rarely adds anything to an ordinary piece of writing.
Prompt chaining
With prompt chaining you split a complex task into a sequence of prompts, and the output of the first prompt becomes the input for the second. Prompt chaining works for tasks that are too big for one prompt: research that first needs keywords, then a brief, then a draft text.
Keeping the steps separate gives you more control at each step and lets you adjust along the way. Prompt chaining is also the basic structure of almost every automated AI workflow: in production you rarely see one gigantic prompt, but almost always a chain of smaller steps with validation in between.
Giving context to an LLM
Giving context to an LLM in your prompt is not a technique in itself but a skill that strengthens all the techniques above. Giving context to an LLM means supplying documents, providing background information and making earlier agreements explicit.
The richer the context, the more accurate the output. Giving context is the skill professionals neglect most often, even though its effect is immediately noticeable. A prompt without context forces the model into general assumptions; a prompt with context lets the model tailor the output to your situation.
Prompting for different AI models
Prompting does not work identically across AI models. A prompt that gives excellent output in ChatGPT can lead to mixed results in Claude, and the other way round. The difference lies in how the models were trained, which trade-offs their makers chose and which style of instruction the model expects.
Which models exist and what each one does best is set out in the overview of well-known AI models.
In my own work I find that Claude sticks to instructions more strictly than ChatGPT. This stands out most in complex multi-stage tasks and in system prompts for API work. Where ChatGPT tends to go off on creative tangents, Claude stays closer to what you ask.
That makes Claude more pleasant for production work where predictability matters more than creativity, and ChatGPT more pleasant for brainstorming and exploration.
With reasoning models, such as Claude Opus 5.5 and ChatGPT at a higher reasoning level, different rules apply to prompting. These models work step by step internally, which makes classic chain-of-thought instructions redundant. Prompts that are too long or too elaborate can actually throw reasoning models off; short, clear instructions work better here than lengthy explanations up front.
When I moved from Claude Opus 4.6 to 4.7, I found that existing prompts had to be retuned to keep the same quality of output. A new model version is not automatically a better model version for your existing workflow.
Common prompt engineering mistakes
Common prompting mistakes are usually not one-off slips but patterns that professionals repeat until they consciously stop. I come across five of them regularly.
The first mistake is giving vague instructions and hoping the model will make something usable of them. “Write something about X” produces generic output by definition. If you make this mistake, you leave the thinking to the model, and that does not make the result any better.
A second pitfall is not specifying a format for the output. Without one, the model gives whatever is statistically most likely to fit, which rarely matches exactly what you want. A clear instruction about the format often saves more than two rounds of correction afterwards.
On top of that, many users ask for things the model cannot do, such as real-time information or legal advice that carries liability. AI models also produce text on subjects they cannot judge, and that text sounds convincing. That is exactly where things go wrong.
During the period when I worked intensively with ChatGPT, I saw a recurring pattern. In keyword research, the model added search volumes without being asked, even though ChatGPT has no connection to Ahrefs, SEMrush or any other keyword tool. These were pure hallucinations, plausible to anyone without keyword data of their own.
The fourth mistake is the most dangerous: treating prompting like a Google search. If you use ChatGPT or Claude as a search engine and copy the output without checking it, sooner or later you will be caught out by an AI hallucination: a fabricated fact that sounds convincing. AI models generate text; they do not look up facts. The difference feels subtle, but it changes how you should treat every answer.
The fifth mistake is pasting confidential information into a prompt: client names, national identification numbers, medical data or a contract containing personal data. Under the GDPR, your organisation remains responsible for what happens to that data, including in an AI tool. In August 2024 the Dutch Data Protection Authority (Autoriteit Persoonsgegevens, AP) reported that it had received several data breach notifications because employees had entered personal data into an AI chatbot. Whether ChatGPT is safe for business use depends mainly on your plan and on the setting for training on your data.
Learning better prompting techniques does not make these mistakes disappear. They disappear when you place prompting in a way of working where checking is a fixed step. That change of mindset matters more than any single technique.
Prompting in a workflow you can check
Building prompting into a workflow you can check means more than writing prompts. You also decide how you will verify that the output is right before it goes any further. This is where many professionals still leave a gap. The prompt is good, the output looks good, and it ends up unchecked on LinkedIn or in a client proposal, until someone notices that a statistic is wrong or a quote was never actually said.
A workable prompting workflow has three fixed steps. Step one is the prompt itself, built from the five building blocks. Step two is a first check by you: are the facts correct, does the tone fit, is anything missing. Step three is a second check by another person or by a different AI.
From my years at major banks I bring the four-eyes principle into this: nothing goes out without a second check. In my own workflow, a test run on a small sample before bulk processing catches most problems before they cause any harm.
For recurring tasks it pays to write the workflow down. The same prompt for keyword research, content briefs or client communication does not need to be reinvented each time. By saving your best prompts, documenting when they work and when they do not, and building them into a fixed way of working, you save time and reduce the chance of errors. A collection of tried and tested ChatGPT prompts for each task is a practical starting point, provided you test each prompt on your own work first. Teams that want prompting to be a shared way of working, rather than something each person does differently, record their prompts and checkpoints in a shared prompt library. Once the whole team applies that way of working in the same way, you are moving into AI adoption.
Not every task needs the same intensity of checking. An internal note has a different risk profile from a published article or client advice: decide in advance which checks are needed for which type of output, and give high-stakes output more checks than low-stakes output. In practice that distinction is still too often ignored, and everything gets the same check, which is too light.
If you are just starting out, you will get most from the five building blocks and basic techniques such as zero-shot, few-shot and role prompting. If you already work with AI regularly, focus on the workflow: how you check output, document prompts that work and make prompting reproducible rather than one-off.
The ChatGPT Course builds this way of working step by step, with practical assignments and a work portfolio you keep using afterwards. If you prefer to work with Claude, you follow the same approach in the Claude Course. LearnLLM was set up to teach professionals exactly this kind of repeatable AI way of working, rather than isolated tricks that are soon forgotten.
