What is prompt engineering—definition and best practices?

Evaluating Prompt Quality & Iterative Prompt Tuning#

Getting a good prompt is partly art, partly measurement:

• Evaluation metrics: For a prompt, track correctness, consistency (variance over multiple runs), and latency (cost). Use human or automated checkers to score outputs.

• A/B comparisons: Use two prompt variants and compare output quality on the same inputs.

• Prompt tuning / automated search: For high-volume tasks, prompt engineers sometimes train soft prompt embeddings or automatically search prompt permutations by performance.

• Prompt libraries / templates: Build reusable templates in a library (e.g. “summarize→bullet points,” “question answerer,” “translate + tone”) and version them.

By treating prompt engineering as an engineering process, you improve dependability and reuse.

Key takeaways:

• Prompt engineering is crafting effective instructions for AI language models to generate desired outputs.

• Prompts are the instructions given to LLMs to guide their responses.

• Prompts can be used for various tasks, including summarization, code generation, translation, and reasoning.

• Best practices for prompt engineering include being specific, clear, and creative.

• Principles of prompt engineering emphasize understanding the LLM’s capabilities, tailoring prompts to the desired output, and iterating on prompts.

• Tips and tricks include using natural language, providing context, and breaking down complex prompts.

• Essential prompt keywords often relate to the desired output, such as “write,” “generate,” or “explain.”

• Common pitfalls in prompting include ambiguity, over-reliance on keywords, and neglecting the LLM’s limitations.

LLMs are booming. They’re everywhere, from writing emails to generating code. But how do you talk to them? It’s all about the prompt. Think of it as giving an AI a super-specific Google search query. The more detailed and clear your prompt, the better the response. For example, instead of saying, “Tell me about dogs,” you may ask, “What’s the best breed for a first-time owner living in a small apartment?” Intrigued? Let’s dive into the art of prompting.

Prompt engineering is designing high-quality prompts that guide machine learning models to produce accurate outputs. It involves choosing the correct type of prompts, optimizing their length and structure, and determining their order and relevance to the task.

"Prompt engineering is the art of communicating eloquently to an AI."

-Greg Brockman on X

Prompt engineering is highly valuable for individuals in various roles, including data scientists, marketers, educators, journalists, writers, business leaders, and entrepreneurs. This blog will introduce prompts and their types, and offer best practices to produce high-quality prompts with precise and useful outputs.

Before digging deeper, let's first discuss the concept of generative AI and large language models.

Generative AI#

Generative AI is a type of an artificial intelligence technology that can generate various content types, including text, images, audio, videos, and synthetic data. Unlike other types of AI that rely on pre-existing data to make decisions, generative AI learns patterns and relationships in the input data, and uses those to generate new and unique output data. 

Large language models (LLMs)#

Large language models (LLMs) are machine learning models that can generate natural language text with impressive quality and fluency. They are trained on massive text datasets using deep neural network architectures such as transformers, and can learn to predict the probability distribution of words in a text sequence. We can also tailor LLMs to your specific needs by training them on custom datasets.

ChatGPT: a type of LLM#

Curious about the power of ChatGPT and other LLMs? Dive into these exciting courses to learn how to harness their potential.

What is a prompt?#

The prompt can take various forms, such as a sentence, a question, a paragraph, or an instruction. The following are a few examples of prompts used for LLMs:

• What is the capital city of the United States of America?

• List the top five most played sports.

• Explain the difference between “affect” and “effect.”

Types of prompts#

Prompts can be used to achieve various kinds of tasks. Understanding how to write prompts to attain the desired result effectively is essential. Prompts are useful in the following tasks, including but not limited to:

1. Text summarization

2. Information extraction

3. Question and answer systems

4. Text classification

5. Translation

6. Code generation

7. Reasoning

Now, let's explore the specifics of the best practices regarding crafting prompts.

Zero-shot, Few-shot and Chain-of-Thought Prompting#

When working with large language models (LLMs), the prompting style can drastically affect performance.

Zero-shot prompting#

You ask for the output without giving examples — ideal for simple tasks when the model already knows the pattern.

Example:

Few-shot prompting#

You include one or more input–output examples in the prompt to show the model how to perform the task.

Example:

Chain-of-Thought prompting#

You instruct the model to “think step by step” before producing the final answer — powerful for reasoning or multi-step problems.

Example:

The model reasons: subtract 5 → 15, divide → 5, then returns “x = 5.”

By combining few-shot and chain-of-thought prompting, prompt engineers can often coax correct reasoning in otherwise tricky tasks.
These techniques are core to advanced prompt engineering.

Prompt Chaining and Decomposition#

Large, complex tasks often overwhelm models when asked in one shot. Prompt chaining splits them:

1. Decompose the task: Break down what you want into subcomponents.

2. Sequential prompts: Prompt for one part, feed the answer into the next.

3. Context retention: Always pass relevant history or intermediate results.

Example for a blog post generation:

• Prompt 1: “Generate a table of contents for a blog on ‘Time management.’”

• Prompt 2: “Write section 1: Introduction using that TOC entry.”

• Prompt 3: “Write section 2 based on the next heading,” etc.

You can also chain verification prompts, e.g. “Check if section 2 has factual errors. If yes, rewrite.” Decomposed prompts often yield more coherent, accurate outputs than monolithic ones.

Best practices#

The principles of prompt engineering#

Let's look at a few principles of prompt engineering with examples since they provide useful guidelines for creating effective prompts that ensure accurate results.

Simplicity #

Simplicity is an essential factor to consider when crafting prompts for natural language processing models. The prompts should be concise, clear, and easy to understand for both the model and the end user. Using overly complex language or providing unnecessary information can confuse the model and lead to inaccurate results.

For example, the following prompt may be too wordy and convoluted for the model to accurately understand and generate the desired output.

In contrast, the following prompt is simple and contains only the necessary information to guide the model toward the desired output:

Specificity #

Specificity is an essential aspect of prompt engineering in natural language processing since it ensures that the generated output is relevant and accurate. When crafting prompts, it is crucial to be specific about the desired output, task, or objective. Generic prompts may not guide the model enough to generate accurate results. 

For example, the following prompt is too general and could result in a wide range of output descriptions that may not be relevant to the user's needs.

However, consider a more specific prompt that provides clear guidance to the model and helps ensure the generated output is relevant and accurate.

By providing specific details in the prompt, we can help the model to focus on the relevant aspects of the task and improve the accuracy of its results.

Tips and tricks for writing good prompts#

Here are some tips and tricks to help you write great prompts:

• Use natural language: Avoid overly formal or technical language. Use relevant keywords to help the LLM understand your query.

• Be creative and experiment: Don’t be afraid to experiment with different prompt formats and styles. Assign the LLM a specific role or persona to guide its response.

• Break down complex prompts: If your prompt is complex, break it into smaller, more manageable parts. Also, provide intermediate steps or questions to guide the LLM through the process.

• Provide examples:  Give the LLM examples of what you’re looking for to help it understand your intent.

• Iterate and refine: Getting the desired response may take a few iterations. Analyze the LLM’s output to identify areas for improvement in your prompts.

Essential prompt keywords#

Essential prompt keywords are specific words or phrases that convey the intended meaning and guide the natural language processing model toward generating the desired output. Including relevant keywords in prompts ensures the model understands the task or objective and produces accurate results. 

For example, consider a prompt like “Summarize the main points of a news article about climate change.” In this prompt, the essential keywords are “summarize,” “news article,” and “climate change.” These keywords guide the model on what task to perform, what type of input data to expect, and what topic to focus on. 

Other examples of essential prompt keywords include verbs that specify the desired action, such as “generate,” “classify,” or “translate,” as well as specific nouns that describe the input data, such as “image,” “text,” or “audio.” Including essential prompt keywords helps to ensure that the natural language processing model produces accurate and relevant results that meet the user's needs. Here's the list you need to get familiar with:

Ready to level up your professional game? Learn essential productivity prompts in this must-read blog:

https://www.educative.io/blog/how-to-write-productivity-prompts-for-professional-roles

What can go wrong while prompting?#

Various factors can influence the accuracy and relevance of results generated by natural language processing models. Here are a few examples:

Ambiguity

Write about the benefits of using social media.

Bias

Prove that climate change is a hoax.

Insufficient context

What is the best restaurant in town?

Too specific

Write a story about a girl named Sarah who goes on a picnic.

Prompt Safety, Guardrails and Format Constraints#

To make reliable prompt engineering, you must apply guardrails:

• Format constraints: Instruct the model to return in JSON, markdown, bullet lists, or strict templates. Helps with downstream parsing. Example: “Return answer as JSON with fields: { ‘summary’: ..., ‘key_points’: [...] }.”

• Self-critique prompts: After generating an answer, ask the model to assess or improve it. Example: “Here’s your draft. Rate it 1–10 for clarity. Rewrite if < 7.”

• Content restrictions: Include constraints to avoid disallowed content. Example: “Generate no personal data. Avoid political bias. Assume content is suitable for general audiences.”

• Tool verification: When prompting models that may call external APIs or tools, instruct them how to handle failures and fallback gracefully.

Without guardrails, hallucinations, harmful outputs, or misformatted text are more likely. Robust prompt engineering demands these safety patterns.

Limitations#

LLMs are incredibly powerful tools for generating text that is often indistinguishable from human writing. However, despite their impressive capabilities, these models have certain limitations. A few of them are detailed in the following paragraphs.

Citing references#

Citing references is a crucial aspect of many types of writing, including academic and scientific publications. However, LLMs can sometimes fail to provide proper attribution, leading to issues with accuracy and credibility. For example, a language model may generate the following sentence:

Solving math problems#

Solving math problems is another area where large language models can fall short. While these models are excellent at generating text, they are not designed to handle complex mathematical equations or operations. For example, an LLM might not be able to generate an accurate answer if it’s asked to solve the following equation:

Hallucination#

LLMs can sometimes generate outputs that are not grounded in reality. Hallucination can occur when the model generates text based on incomplete or incorrect information. For example, a model may generate the following sentence:

In conclusion, while LLMs are impressive and powerful tools, it's essential to be aware of their limitations. Careful consideration and appropriate use of these models can help mitigate these limitations and maximize their potential benefits.

If you’re interested in learning more about prompt engineering, look no further! Check out the exciting new courses available on the Educative platform:

1. All You Need to Know About Prompt Engineering

2. Empowering Solopreneurs in the Enterprise Landscape with ChatGPT