How to Write a Good Prompt?

A prompt is essentially nothing more than an instruction to an AI model. It can be a simple question, a complex work assignment, or a multi-part briefing with context, examples, and a desired format. The AI model takes this prompt as input and generates its response based on it.

The crucial point: AI models are not mind readers. They interpret exactly what you write, not what you mean. The latest generation of models like Claude 4.x in particular follows instructions far more literally than their predecessors. While older models often generously interpreted and supplemented vague requests, current models do exactly what you ask them to do, nothing more and nothing less. (Anthropic, Claude 4.x Prompting Best Practices, docs.anthropic.com, 2026 / DreamHost Blog, We Tested 25 Popular Claude Prompt Techniques, December 2025)

This means: the more precise your prompt, the better the result. And that’s exactly why it’s worth understanding the fundamental principles of good prompt engineering.

The most important rule is also the simplest: tell the AI exactly what you want. The more specific your instruction, the less the model has to guess and the better the result will be. Instead of “Explain climate change to me,” it’s better to write: “Write a three-page summary of climate change for a specialist audience in the environmental sector, focusing on the economic consequences for Europe.” Clarity is the cornerstone of every good prompt, something confirmed by both OpenAI and Anthropic in their official guides. (OpenAI, Prompt Engineering Guide, platform.openai.com, 2025 / Anthropic, Be Clear and Direct, docs.anthropic.com, 2026)

One of the most powerful techniques is to assign the model a role. When you write “You are an experienced marketing consultant with 15 years of experience in the B2B sector,” it activates different knowledge patterns than a general request. The model then draws on texts and data that match this role and delivers correspondingly more specific answers. This technique is also known as “persona prompting” and is particularly effective when you need specialized expertise or a certain tone of voice. (Anthropic, Give Claude a Role, docs.anthropic.com, 2026 / Lakera, The Ultimate Guide to Prompt Engineering, 2026)

Should the response come as flowing text, a table, a bulleted list, or JSON? Should it be three sentences or three pages long? Format specifications like these help the model enormously. A prompt like “Create a comparison of the three providers as a table with columns: Name, Price, Advantages, Disadvantages” delivers far more structured results than “Compare the three providers.” According to DigitalOcean, clear format definition is among the most effective best practices in prompt engineering. (DigitalOcean, Prompt Engineering Best Practices, 2025 / OpenAI, Best Practices for Prompt Engineering, help.openai.com)

The more relevant context you include, the better the AI can place your request in perspective. This can be background information about your company, your target audience, previous decisions, or relevant data. Anthropic explicitly recommends in its documentation placing longer context documents at the beginning of the prompt and formulating the actual question afterward. Tests have shown that this ordering can improve response quality by up to 30 percent. (Anthropic, Long Context Tips, docs.anthropic.com, 2026 / DreamHost Blog, Claude Prompt Engineering, December 2025)

One of the most effective methods for showing a model what you expect is providing examples. If you want the AI to summarize customer feedback in a specific style, for instance, include one or two examples in your prompt. The model recognizes the pattern and applies it to new inputs. In research, this technique is called “few-shot prompting,” as opposed to “zero-shot prompting,” where no examples are provided. Studies show that few-shot prompts can improve accuracy over zero-shot by an average of 12 to 13 percent. (Anthropic, Use Examples / Multishot Prompting, docs.anthropic.com, 2026 / ScienceDirect, Applying LLMs and Chain-of-Thought for Automatic Scoring, 2024)

For complex tasks, it helps to explicitly ask the model to proceed step by step. This technique is called “Chain-of-Thought Prompting” (CoT) and was first systematically studied by researchers led by Jason Wei at Google in 2022. The results were impressive: for mathematical and logical tasks, accuracy improved significantly when the model was asked to reveal its thinking steps.

However, more recent research from the Wharton School paints a more nuanced picture: with the very latest models that have built-in reasoning (such as o3-mini or Claude with Extended Thinking), additional CoT prompting yields only marginal improvements of 2 to 3 percent, because these models already think step by step on their own. For older or simpler models, CoT remains a highly effective technique. (Wei et al., Chain-of-Thought Prompting Elicits Reasoning in LLMs, Google Research, 2022 / Meincke et al., The Decreasing Value of CoT, Wharton School, June 2025)

The perfect prompt rarely emerges on the first attempt. Prompt engineering is an iterative process: you start with an initial draft, evaluate the result, and adjust the prompt step by step. Often, small changes are enough, such as an additional sentence of context, a clearer format specification, or a concrete example, to significantly improve output quality. Both OpenAI and Anthropic emphasize in their documentation that iterative testing is the key to consistent results. (OpenAI, Prompt Engineering, platform.openai.com, 2025 / DigitalOcean, Prompt Engineering Best Practices, 2025)

Mistake 1: Being too vague. “Write me something about marketing” is not a good prompt. Better: “Write a 500-word blog article about three trends in B2B content marketing in 2026, focusing on the German market.”

Mistake 2: Too many tasks at once. If you want a single prompt to research, analyze, summarize, and format all at the same time, quality suffers. Break complex tasks into multiple steps instead.

Mistake 3: Not providing context. The AI doesn’t know your company, your industry, or your target audience unless you tell it. Without context, the response stays generic.

Mistake 4: Negations instead of positive instructions. “Don’t write a boring text” doesn’t tell the AI what it should do instead. Better: “Write a lively, concrete text with practical examples.” OpenAI explicitly recommends using positive instructions in its best practices. (OpenAI, Best Practices for Prompt Engineering, help.openai.com)

Mistake 5: Not reviewing and iterating on results. Anyone who accepts the first output without questioning or refining it is leaving potential on the table. The best results come from dialogue with the AI, not from a single prompt.

A fascinating insight from current research: the more powerful models become, the simpler prompts can be. Researchers refer to the so-called “Prompting Inversion,” a phenomenon where elaborate, highly structured prompts actually deliver worse results on very large models than simple, natural instructions. With GPT-4o, complex prompts still improved accuracy from 93 to 97 percent. With GPT-5, however, the same prompts led to a decline to 94 percent, while simple chain-of-thought prompts achieved 96 percent. (Khan, The Prompting Inversion, arXiv, October 2025)

This doesn’t mean prompt engineering is becoming obsolete. On the contrary: the ability to choose the right strategy for the right model and the right task is becoming increasingly important. The core principles (clarity, context, examples, and iteration) remain universally valid. What’s changing is the balance between structure and trust: with newer models, you can leave more to the AI and should micromanage less.

Using the chatbot ChatGPT as an example, Alexander Führen explains tips and tricks for more efficient use:

Do you need to know how to code to write good prompts? No, absolutely not. At its core, prompt engineering is a linguistic skill, not a technical one. It’s about communicating clearly and in a structured way. Programming knowledge can help with technical applications, but it’s not required for the vast majority of use cases.

What’s the best AI model for beginners? ChatGPT and Claude are equally well suited for getting started. Both offer free plans, respond well to simple prompts, and quickly deliver usable results. If you need to process longer texts, Claude’s large context window is particularly beneficial.

How long should a good prompt be? As long as necessary, as short as possible. For simple tasks, one or two sentences are enough. For complex projects, prompts can span an entire paragraph. What matters is not the length, but the clarity and completeness of the instruction.

Do the same prompts work with every model? Not always. The fundamental principles like clarity, context, and examples apply universally. However, models respond differently to formatting, structure, and the level of detail. A prompt that works perfectly with Claude may need to be slightly adjusted for ChatGPT.

What’s the most common prompting mistake? Being too vague. Most poor AI results can be traced back to imprecise instructions. Anyone who makes a habit of defining the task, context, format, and target audience will see immediate and significant improvement in their results.

Will prompt engineering still be relevant in the future? Yes, but it will evolve. While highly complex prompting techniques are becoming less necessary with the latest models, the core competency, clear communication with AI systems, remains one of the most important skills of the coming years. The models are getting better, but the quality of the input still determines the quality of the output.

(Anthropic, Prompt Engineering Interactive Tutorial, GitHub / Learn Prompting, learnprompting.org / OpenAI, Prompt Engineering Guide, 2025)

The ability to write good prompts is no longer a niche skill in 2026. It’s a fundamental competency for anyone working with AI. The good news: you don’t need a technical background or expensive courses. Anyone who internalizes the core principles (clarity, context, role, format, examples, and iteration) will achieve better results with every prompt.

As with most skills: practice makes perfect. Start small, experiment with different techniques, and gradually build up a collection of proven prompt templates. The investment is worth it, because the quality of your prompts directly determines the quality of your AI-powered work.

Want to use AI tools professionally in your company? We at ThatWorksMedia are happy to help, from prompt strategy to e-learning production to AI-powered content creation. Contact: thatworksmedia@gmail.com