What is prompting? Meaning, prompt engineering and techniques

In a short time, prompting has become an everyday word. Look it up and you will find dictionaries that explain it and courses that go straight to tips. Here you first get the meaning of prompting, and then the step to prompt engineering: designing instructions that work every time, including on hard tasks and in the hands of a colleague. We list the techniques that have a real research basis, from few-shot to self-checking, with the papers included. And we set side by side what OpenAI, Anthropic and Google write about it themselves today, because with reasoning models part of those techniques works differently than with the models that came before.

What does prompting mean?

Prompting is giving instructions to an AI model in plain language: the question, the context, any examples and the desired form of the answer. AI prompting is the same skill, applied to language models such as ChatGPT, Claude or Gemini. If you approach prompting systematically, with testing and reusable templates, you are doing prompt engineering.

Google describes prompt design as creating requests in natural language that get a language model to give accurate, high-quality answers. We explain what a prompt is and how to write a good one in what is a prompt?. Here it is about the verb, and about how you turn that instructing into a craft.

Outside AI, the word has an older meaning. In psychotherapy, according to the APA Dictionary of Psychology, prompting is the therapist suggesting topics to encourage the client to discuss certain issues, for example by reminding the client of previously discussed material or finishing a sentence for the client. Speech therapy also has the PROMPT method, in which, according to the PROMPT Institute, the speech therapist uses their hands to guide the jaw, lips and tongue to form words. From here on, this article is only about AI.

In Flanders, prompting is no longer a niche. According to imec.digimeter 2025, 43% of Flemish people aged 18 and over actively use generative AI, and 51% of working people do so at least monthly for work. How good their instructions are determines what they get out of the same tools. ClickForest teaches teams of SMEs in Flanders prompting and prompt engineering in its AI training courses, on the tasks they do every day.

What is prompt engineering?

Prompt engineering is the systematic design, testing and refinement of prompts so that an AI model reliably delivers the desired output. The difference with everyday prompting lies in that word systematic: you build a reusable instruction, test it on multiple cases, look at where it fails and sharpen it until the quality is consistent.

This is how OpenAI itself describes it:

“Prompt engineering is the process of writing effective instructions for a model, such that it consistently generates content that meets your requirements.”

— OpenAI, prompt engineering guide in the API documentation (accessed September 2026)

How you word an instruction largely determines what a language model delivers; that is why OpenAI, Anthropic and Google each publish their own prompting guide. For the daily basics (the Role + Goal + Context + Output formula and ready-made example prompts) we refer to our beginner’s guide on prompt writing; this article starts where that guide stops. A concise definition is also in our glossary.

What is the difference between prompting and prompt engineering?

Prompting is asking a question; prompt engineering is writing a recipe. Whoever prompts wants a good answer now. Whoever does prompt engineering wants an instruction that next week, used by a colleague, on a different case, still hits the same quality level. That requires a different way of working: designing, testing on multiple examples, and measuring.

In practice you see the difference in what happens to a prompt after it has worked once. An engineered prompt becomes a template with fill-in fields that you use again and again. Before you share it, you put it next to five to ten real cases and judge the output systematically, because one successful attempt says little. What works, you write down, so that a colleague can build on it later. That is how prompt engineering becomes the bridge between loose AI use and the structured AI marketing way of working a team can build on. If you like a fixed template for everyday prompting, you will find one in the RACE model for ChatGPT prompts; the techniques below start where such fill-in models stop.

Which prompt engineering techniques should you know?

Four techniques cover the vast majority of practice: giving examples (few-shot), step-by-step reasoning (chain-of-thought), chaining prompts (chaining) and letting the model check its own work. All four have been studied in scientific papers. How you use them does differ per model: the newest reasoning models think for themselves and no longer need part of those instructions.

Technique	What you do	When to use it
Few-shot (giving examples)	You show 3 to 5 examples of input plus desired output before the real question	Style, format or judgement that is hard to describe but easy to show
Chain-of-thought (step by step)	You ask the model to reason first and only then answer	With models without built-in reasoning; reasoning models do this themselves
Prompt chaining (linking steps)	You cut a big task into steps, each with its own prompt	Long workflows where you want to check the intermediate results, such as research, outline, text and review
Self-checking	You have the model test its output against a checklist or criteria	Quality control in recurring production, before the human final check

Chaining and self-checking often go together. According to Anthropic, the most common chaining pattern is self-correction: have a first draft made, have it reviewed against criteria and then have it refined. That also works within a single prompt. Anthropic recommends asking the model to verify its answer against your test criteria before finishing, which according to the documentation catches errors reliably, especially in code and maths.

The scientific roots are worth knowing. Few-shot learning was put on the map by the GPT-3 paper by Brown and colleagues (2020, “Language Models are Few-Shot Learners”), chain-of-thought by Wei and colleagues (2022), who showed that eliciting reasoning steps improves performance on arithmetic, commonsense and symbolic reasoning tasks. Wu and colleagues described chaining prompts in 2021 (“AI Chains”): the output of one step becomes the input of the next. Madaan and colleagues studied having the model assess and improve its own output in 2023 (“Self-Refine”). The papers explain why the techniques work: you steer the way the model arrives at its answer. To use them, you do not need to have read them.

“Experiments on three large language models show that chain of thought prompting improves performance on a range of arithmetic, commonsense, and symbolic reasoning tasks.”

— Jason Wei and colleagues, abstract of the chain-of-thought paper (January 2022)

What is the difference between zero-shot, one-shot and few-shot prompting?

The difference lies in the number of examples you include in the prompt. With zero-shot you give only an instruction, without an example. With one-shot you add exactly one example, with few-shot a handful. The GPT-3 paper of 2020 described them as three separate settings. More examples steer form and tone more strongly, and also make your prompt longer.

Setting	What you include	When it is useful
Zero-shot	Only an instruction in plain language, no example	Simple tasks, and the first attempt with a reasoning model
One-shot	The instruction plus one example of input and desired output	When one good example already makes the desired form clear
Few-shot	The instruction plus a handful of examples; Anthropic recommends 3 to 5	Style, tone or judgement that is easier to show than to describe

The GPT-3 paper describes one-shot as the same setting as few-shot, with only one demonstration next to the description of the task, and zero-shot as a setting in which the model is only given an instruction in natural language. Google uses the same division in its Gemini documentation: prompts with a few examples are called few-shot, prompts without examples zero-shot. On the value of examples, Anthropic, the company behind Claude, is outspoken:

“Examples are one of the most reliable ways to steer Claude’s output format, tone, and structure. A few well-crafted examples (known as few-shot or multishot prompting) improve accuracy and consistency.”

— Anthropic, Prompting best practices in the Claude documentation (accessed September 2026)

Anthropic also asks that those examples are close to the real task, cover edge cases and vary enough so the model does not pick up unintended patterns, and that they are clearly separated from the instructions. Which setting works best depends on the model. That is what the next question is about.

How do you prompt a reasoning model?

Give a reasoning model a clear goal and the necessary context, and leave the thinking to the model. Such models first work out a line of reasoning internally before they answer. OpenAI therefore advises against still asking them to think step by step, and recommends trying without examples first. Test for yourself whether examples still add anything with your model.

OpenAI calls such models reasoning models, Anthropic and Google speak of thinking. That changes part of the techniques above. On step-by-step thinking the three vendors are close to each other; on examples they differ:

Vendor	Asking for step-by-step thinking	Including examples
OpenAI (guide for reasoning models)	Advised against: “Avoid chain-of-thought prompts”, because those models already reason internally	Without first: “Try zero shot first, then few shot if needed”
Anthropic	A general instruction rather than a written-out step-by-step plan; manual step-by-step reasoning is a fallback when thinking is off	3 to 5 examples, and those work together with thinking
Google (Gemini)	Usually not needed: according to Google it is “generally not necessary” to have the model write out its reasoning steps. For heavy reasoning tasks, a simple request like “Think very hard before answering” can help, at the cost of extra thinking tokens	Always: “We recommend to always include few-shot examples in your prompts”

Anthropic puts it most sharply: “A prompt like ‘think thoroughly’ often produces better reasoning than a hand-written step-by-step plan.” For its recent Gemini models, Google also asks you to phrase prompts directly and in a structured way, with the task and the constraints clearly defined. OpenAI sums up the difference between the two kinds of models with a comparison from the workplace:

“A reasoning model is like a senior co-worker. You can give them a goal to achieve and trust them to work out the details. A GPT model is like a junior coworker. They’ll perform best with explicit instructions to create a specific output.”

— OpenAI, prompt engineering guide in the API documentation (accessed September 2026)

Because the vendors disagree about examples, the mini test set from the following sections becomes more important: only by testing do you see whether an example or a step-by-step plan still adds something with your model. We compare which assistant suits which task in ChatGPT vs Claude vs Perplexity.

How do you build a prompt library for your team?

For an SME, a prompt library is the biggest lever of prompt engineering: a shared collection of tested prompt templates per task, with version control in its simplest form, namely who changed what and why. That way, what one employee tries out benefits the whole team, and it is kept when that employee is busy with something else.

Start with the three most recurring tasks, give each template an owner and agree on a fixed format: name, goal, the template itself, a good example of the output, the model it was tested on and the last change. Store it centrally. A shared document is enough, specialised tools come later. Plan a short revision every quarter, because models change and templates age with them. You notice the difference most with new colleagues: they start right away with the team’s templates. For the broader tool choice around such a way of working, our guide to the best AI tools for marketers helps.

Want to practise this with your team? ClickForest trains teams of SMEs in Flanders in prompting and prompt engineering on their own tasks, in a session of half a day or a full day, or in a track of several sessions. Take a look at the AI training or book a no-obligation video call.

How do you test whether a prompt is really good?

With a mini test set: put your prompt next to five to ten real, varied cases, including a few tricky ones, and judge the output on fixed criteria such as factual accuracy, tone and usability without rework. Only when the output is consistent across the whole test set do you know that the prompt works and can you share it with your team.

OpenAI explains in its guide why that testing is needed: “Because the content generated from a model is non-deterministic, prompting to get your desired output is a mix of art and science.” The same model can therefore answer the same prompt differently each time.

Keep the testing light, or you will not keep it up. A simple scorecard is enough: good, usable with edits, unusable. Keep the tricky cases as a fixed test set, so every prompt version is measured against the same bar. If the underlying model changes, run the test set again: prompts that worked well for months can respond differently after a model switch. OpenAI itself warns that even different versions within the same model family can give different results, and recommends using fixed model versions and building tests.

When do you build a tested prompt into an automation?

As soon as a prompt scores reliably across your whole test set, you can build it in: into a marketing automation that runs it every time, or into an AI agent that uses it as part of a bigger task. Prompt engineering is thus the quality gate to automation. What does not work consistently with a human involved does not belong in a system without one either.

The order matters: first make the prompt reliable, only then connect it to workflows, and always with a checkpoint wherever mistakes cost money or reputation. ClickForest builds such marketing automations and AI agents for SMEs in Flanders. Our guide on AI agents in marketing shows what becomes possible once that base is in place; which assistant to use for what is in our practical guide to choosing between ChatGPT, Claude and Perplexity.

Is prompt engineer a job, and should you hire one?

The job title exists, but never really took off. According to Fortune, searches for such jobs on Indeed already peaked in April 2023. More and more often, prompt engineering becomes a competence within existing roles: the marketer who builds reliable content prompts, the analyst who designs reasoning prompts for reports. For most SMEs, training is therefore a better route than hiring.

“Prompt engineering as a skill is still definitely a good thing to have, but it’s not an entire title”, Indeed economist Allison Shrivastava told Fortune in May 2025. Those figures and that statement concern the US labour market.

The skill itself is still often missing in Flanders. According to imec.digimeter 2025, 33% of Flemish people do not know how to use generative AI to carry out tasks more efficiently, and only 27% of working people say their employer has a clear policy and guidelines on generative AI. You learn the techniques in this article in a few days, and they pay off most with the people who already know the tasks and the customers. That is why this combination works: teach your own people the skill, with external guidance for the start and the hard cases. ClickForest gives that AI training to SMEs in Flanders, in-company or online.

What should you remember about prompting and prompt engineering?

Prompting is giving instructions to an AI model. Prompt engineering is the deeper layer on top: systematically designing, testing and refining until a prompt is reliable enough to share, repeat or automate. The four core techniques have a serious research basis, but reasoning models call for a lighter hand: fewer step-by-step plans, more goal and context.

For a team, the biggest gain is in the prompt library: shared, tested templates that let knowledge stack up, with a test set you run again at every model switch.

ClickForest teaches teams of SMEs in Flanders prompting and prompt engineering on their own tasks, as part of a hands-on AI training or a broader AI programme. Curious where your team stands? Book a no-obligation video call and we will look together at which tasks qualify first.