Karel from the marketing team taps into his keyboard: "Write me something about our new product for social media." He hits Enter. Out comes a generic, faceless block of text stuffed with clichés and rocket emojis that wouldn't sell water in a desert. Karel sighs that this AI is basically useless and goes off to write it himself.
Two desks away sits Klára. She thinks for a moment first, then sends over this: "You are a senior B2B copywriter. Our new product is X, it solves problem Y for target audience Z. Write me 3 versions of a LinkedIn post. Use the PAS framework (Problem-Agitation-Solution). Tone: professional, but slightly cheeky. No hashtags, no rocket emojis." Thirty seconds later she has a draft at the level of a junior with three years of experience. All it needs is a light polish.
The difference? It's not in the tool. It's in how you write the prompts.
A quality AI prompt is today the difference between using AI as an expensive calculator and using it as a tireless intern who has read the entire internet. According to recent data, companies that have introduced structured prompting see 340% higher ROI on their AI investments.
At LK Media I went through this hell myself. For the first few months with AI, I sat there swearing and rewriting the outputs from scratch. I was convinced the models were simply stupid. Until one day (in a fit of despair) it dawned on me that the stupid thing was my instruction. From the moment I understood that the prompt is the be-all and end-all, the rules of the game changed for us as a company.
The short version for those who don't have time
- 58% of employees already use AI regularly, but only a fraction do it effectively.
- Standardised prompt engineering has been shown to deliver a 340% return on investment.
- The foundation is 5 techniques: role, examples, step-by-step, structure and iteration.
- Below you'll find copy-paste templates for marketing, sales, HR, finance and customer support.
- A good ChatGPT prompt can genuinely cut your work from several hours down to a few minutes.
Why the prompt is the most important thing you'll learn in 2026
I don't want to bore you with a pile of charts, but the numbers here speak pretty clearly. In 2026, AI is already used regularly by 58% of employees in office professions. But using and making the most of are two different things.
PwC recently published a report showing that there's a wage premium of +56% in the market for people with demonstrable AI skills. And the market for "prompt engineering" (which is really just a fancy name for the art of talking to a machine) is growing at 33% CAGR.
Why? Because companies have finally figured out that when they let people type random things into ChatGPT, the success rate for completing a more complex task sits somewhere around 34%. The moment they give them standardised templates and teach them the structure? The success rate flies up to 87%.
A personal example. Writing a complex content strategy for a new client used to take me around 4 hours of focused time. Research, structure, tone of voice, breaking it down by channel. Today? I have a finely tuned mega-prompt. I feed it the data from the kick-off call, the client's URLs and the goals. Five minutes later I have a skeleton in front of me that covers 80% of the work. I spend the remaining 2 hours (yes, I saved almost 3 hours) adding the human "spark", the strategic nuance and the wit.
No average prompt will give you that.
5 prompting techniques everyone needs to know
This isn't magic. It's the psychology of the machine. Here are five basic principles — once you grasp them, your prompts will be on a completely different level.
1. Role prompting (give the AI an identity)
When you don't tell the AI who it is, it averages out the entire internet. And the average is boring. You have to give it a mask.
Example: "You are a senior copywriter with 10 years of experience in e-commerce. Your speciality is writing product descriptions that increase conversions and sound natural."
Why it works: The model instantly trims its "vocabulary" and style down to a pro's level, filters out the academic textbook stuff and focuses on selling.
2. Few-shot (show what you want)
Most people give "zero-shot" prompts (no example). Mistake. Models learn best from patterns. Give the AI three examples of what you like and it will copy that pattern for the fourth.
Example: "Here are 3 sample emails we've sent clients in the past that got a great response: [insert texts]. Write the 4th email on the topic of a spring sale in exactly this tone, length and style."
3. Chain-of-Thought (step by step)
This is an absolute game-changer for more complex tasks. Try adding a sentence to the end of an analytical prompt sometime: "Think step by step." It increases accuracy on logical tasks by 20 to 40%.
Example: "Analyse this Q1 sales data. Think step by step. First identify the best-selling product, then find the days with the biggest drop, and finally propose a hypothesis for why the drop happened."
4. Structured output (ask for a clear format)
Stop reading paragraphs of text when you need data. Spell out exactly how you want the output to look. Roughly 70% of enterprise companies now require this in their internal processes.
Example: "Give me the output exclusively as a Markdown table. The columns will be: Competitor name, Estimated price, Strength, Weakness."
5. Iteration (it's a dialogue, not Google)
The biggest misconception? People think a prompt is a finished command. No. It's a conversation. Never expect a miracle on the first attempt. (Actually, when it does come out right the first time, I tend to get nervous about what I missed.)
Example: "That's a good starting point, but it sounds too corporate. Rewrite it as if you were explaining it to a friend over a beer. Cut it in half and drop the word 'innovative'."
Which AI tool for what (and how to prompt in it)
Not all artificial intelligence is created equal. It really no longer holds true that "I'll open ChatGPT and solve everything." Sure, you can do it that way, but it's like slicing bread with a hacksaw. It works, but you'll struggle.
OpenAI's ChatGPT is still that universal Swiss army knife. It's great for creative tasks, brainstorming and, thanks to Custom Instructions (where you set out who you are and how you want it to reply), it saves you a lot of time. On top of that, the GPT Store is full of ready-made assistants. When I don't know where to turn, that's where I go.
But then there's Anthropic's Claude. This is my personal favourite for the heavy lifting. Claude has a Projects feature where you upload your own documents, and it has so-called Artifacts — when you're coding or doing web design, it visualises the code for you in a side panel right away. Claude is brilliant at long texts and analysis. It doesn't write as "robotically" as ChatGPT. A crucial tip for Claude: it loves XML tags. When you separate your instructions in the prompt using <context>, <task> and <format>, it'll purr like a cat.
If you're on Google Workspace, you've probably already run into Gemini. It sometimes has a slight tendency to hallucinate (make things up), but its integration into Docs and Sheets is simply convenient. An excellent tool if you need to pull data from your own emails or company documents on Drive.
And finally, Perplexity. It's not so much a text generator as an AI search engine on steroids. I use it exclusively for research and fact-checking. Unlike the others, it gives you a link to the source for every fact, so you can verify that it really didn't just pull it out of thin air.
Copy-paste prompt templates for 5 company departments
Theory is fine, but you want results. Here are some concrete templates. Take them, replace the text in the square brackets and watch what happens.
MARKETING:
Template for a sales email (follow-up)
> "You are an expert in B2B email marketing. Your task is to write a follow-up email for a prospect who downloaded our e-book on [e-book topic] 3 days ago.
> Goal: get a reply and arrange a 15-minute call.
> Style: friendly, concise, not pushy.
> Structure: 1. Reference the downloaded material (add one extra useful idea). 2. A question about their current biggest challenge with [problem area]. 3. A soft CTA to a call.
> Write 2 subject-line variants and the email itself. Maximum 120 words."
Template for a social media post
> "You are a social media copywriter. Write a post for [LinkedIn/Instagram] about our new blog article, titled: [article title].
> Main idea of the article: [1 sentence on what it's about].
> Format requirements:
> 1. A hook (first sentence) that sparks curiosity or challenges a common myth.
> 2. A quick summary of the 3 main points using bullet points.
> 3. A clear CTA to click the link in the comments.
> Tone: authentic, energetic, no corporate jargon."
SALES:
Template for building a proposal
> "Act as a senior sales rep. Prepare a draft sales proposal for the client [company name].
> Context: the client is dealing with a problem around [problem description] and losing [money/time] because of it.
> Our solution: [product/service name], implementation takes [time], the price is [amount].
> The output must include:
> - A summary of their situation (so they see that we understand their pain).
> - How our product solves the situation (focus on benefits, not features).
> - A timeline.
> - The price and next steps.
> Write confidently, but with respect for their situation."
Template for a post-meeting follow-up
> "Write a summary email after a sales meeting.
> Attendees: me and [client name].
> What we discussed: [point 1, point 2, point 3].
> What we agreed on: [outcome].
> Next steps: I'll send [document] by [date], the client will think over [topic] internally.
> Tone: professional, organised. The email must be very easy to scan; use bold for the important dates."
HR (HUMAN RESOURCES):
Template for a job ad
> "You are a modern HR specialist. Write a job ad for the position of [job title].
> Who we are: [brief description of the company and culture].
> What the person will do: [3 main responsibilities].
> What we require: [3-4 key skills].
> What we offer: [salary, benefits].
> Instructions: avoid phrases like 'a dynamic team' or 'we're looking for a superhero'. Write it in a human, honest way and focus on why a smart person would actually enjoy the job. Format the output using headings and bullet points."
Template for screening questions
> "We're preparing interviews for the position of [job title]. Here is the job description: [insert description].
> Suggest 5 behavioural interview questions that will help us find out whether the candidate has real experience with [specific skill, e.g. handling crises with angry clients].
> For each question, add:
> 1. Why we're asking it.
> 2. What signals a good answer (red flags vs. green flags)."
FINANCE:
Template for a management report
> "You are a financial analyst. Here is the raw data on our costs and revenues for last month: [insert data/figures].
> Write a concise executive summary for the CEO.
> Requirements:
> 1. Highlight the 3 most important trends or anomalies.
> 2. Don't overload the text with numbers; use them only to illustrate the main point.
> 3. At the end, add a 'Recommendations' section proposing 2 steps to reduce costs in the category that grew fastest.
> The output must be absolutely to the point and free of any filler."
Template for a forecast (3 scenarios)
> "Based on this historical sales data [insert data], create a rough outlook for the next quarter.
> Work out 3 scenarios:
> - Pessimistic (market decline of 10%)
> - Realistic (continuation of the current trend)
> - Optimistic (growth of 15% thanks to a new campaign).
> Give me the result as a Markdown table so I can copy it straight into a report. Think step by step and explain how you arrived at the figures."
CUSTOMER SERVICE:
Template for a complaint response
> "You are an empathetic customer support specialist. A customer sent us this complaint: [insert email text].
> Write a reply.
> Rules:
> 1. Apologise, but don't be excessively grovelling.
> 2. Show that you understand their frustration (mirroring).
> 3. Offer a concrete solution: [how you'll resolve it, e.g. refund / replacement].
> 4. The tone must be reassuring and professional. Don't shift the blame onto the system or colleagues. We take full responsibility."
Template for a FAQ generator
> "Here is a transcript of emails from our customers over the past week: [insert texts/queries].
> Analyse them and create a FAQ (Frequently Asked Questions) section for our website.
> 1. Identify the 5 most frequent questions.
> 2. For each one, write a concise, clear and understandable answer (max 3 sentences per answer).
> 3. The answers must sound helpful and must include the steps the customer should take."
8 mistakes 90% of people make (and how to stop making them)
When we train companies at our agency, I see it over and over. People slag off AI, but the problem is sitting between the chair and the keyboard. No, really. It's not the machine's fault that you're not getting what you want.
The absolute worst is a vague brief. Remember Karel from the intro? "Write me something…" is a reliable road to hell. Artificial intelligence can't read minds (yet). When you don't clearly say what, you get the average. Hand in hand with that goes an unspecified audience. A text for a bank director has to look different from a text for an eighteen-year-old skater. If you don't put that in the prompt, the AI writes a text for no one.
Then we have a lack of context. Imagine stopping a stranger in the street and saying: "Make me a marketing campaign." They'd probably look at you like you're mad. What the AI does is start making things up. You have to give it the background — who you are, what you sell, why you do it.
Another classic: multiple tasks at once. "Analyse this text for me, find the mistakes, rewrite it, translate it into English and turn it into a tweet." The model gets lost, skips something and botches the rest. Break it up. Step one, step two.
And what winds me up most personally? An unspecified format. People get a mile of text and then manually retype it into a table. Why? Just say straight away: "I want it in bullet points" or "Turn it into JSON".
No iteration is another pain. People enter a prompt, get the first answer, think "ugh" and close the window. No! Tell the model exactly what you don't like about it. "This is too formal, tone it down."
Not using roles we've already covered, but I'll repeat it: without a role the AI simply flounders. And the last thing — ignoring the differences between models. People try to force ChatGPT into a detailed data analysis of a long PDF document, and then they're surprised it hallucinates, instead of using Claude or Gemini for it. Every tool has its strengths.
Advanced techniques for the curious
All right, you've mastered the basics. What next? The world of prompting is moving at a brutal pace. If you want to be a step ahead, you should start looking at system prompts and Custom Instructions. This is essentially an invisible layer of instructions that gets added to every one of your queries. You set it up once and the AI knows forever what your company's tone of communication is and what it must never do.
Then there's RAG (Retrieval-Augmented Generation). It sounds scary, but in practice it means you connect the AI to your own company data (intranet, manuals, order history). Then, when you ask "What are our cancellation terms?", the AI doesn't search Google, it reaches into your PDF and answers exactly according to you.
The absolute hit of 2025 and 2026 is MCP (Model Context Protocol). It's a standard that lets AI models reach directly into your tools. So in Claude you no longer have to copy data from GitHub or Google Drive — via MCP, Claude pulls it in itself, reads it and works with it directly.
(And if you want to go full sci-fi, try multi-agent workflows. You build several AI agents — one writes the text, a second acts as its editor and critiques it, a third verifies the facts. They talk among themselves and spit out only the final, perfect result for you. Oh, and don't forget "temperature" — the parameter you use in an API or better interfaces to control creativity. A temperature of 0.0 means absolute factual precision with no imagination, for finance; a temperature of 0.7 and above is ideal for creative marketing.)
Conclusion
Let's go back to the start, to our imaginary office. Klára is already drinking her coffee and is done, while Karel is still sweating over a blank Word document and cursing modern technology. The best prompts aren't the longest ones. They're the best structured ones.
If you don't want your team to end up like Karel, and you want to actually see the time (and money) saved, at LK Media we offer AI mentoring and in-house consulting. We'll go through your processes and build custom prompts for you.
And if you'd like to get hands-on with something visual and fun, take a look at our Mini-course on AI photo generation. It's a great way to start learning how to communicate with the machine (because with photos you immediately see what your prompt did). Or take it a step further and learn to make videos straight away in our AI video course.
So — what will you type into the chat tomorrow morning?
FAQ
What is prompt engineering?
Put simply, it's the skill of correctly framing an instruction (a prompt) for an AI so that it gives you the most precise, highest-quality and most useful result. It isn't programming, it's more about clear, structured communication.
How do I write a good ChatGPT prompt?
The basic formula is: give it a Role + a clear Task + enough Context + the desired Format. And above all — iterate. If you don't like the result, don't rewrite the whole thing; just tell it in the next step what to fix.
Which AI tools are best for businesses?
It depends on the task. ChatGPT (with a team subscription) is a great all-rounder. Anthropic's Claude is currently king for working with long texts, documents and code. For businesses on Google infrastructure, Gemini Advanced makes a lot of sense.
How much does a prompt engineer cost?
In Western markets, salaries for specialised prompt engineers run extremely high (often above USD 150,000 a year). In the Czech market it's still more of an add-on skill to an existing role (copywriter, analyst, marketer), but one that can lift your value in the job market by tens of percent. For external consulting, expect an hourly rate of roughly CZK 2,000 to 5,000 (≈ €80–200).
How do I get started with AI in a company?
Stop letting people use it for everything and find one specific, annoying, repetitive task (like writing meeting summaries or generating answers to frequent questions). Build one perfect, tested prompt for it, show the team how much time it saves, and only then scale it further.
