Artificial Intelligence (AI) Prompt Writing Guide Ask Us

Introduction

Before using this guide the following information may be useful:

• If you want a general introduction to Artificial Intelligence, check out our guide on Artificial Intelligence for Undergraduate Students.
• Remember, you should never pass off AI generated work as your own and your professor decides if AI use is permitted in their class. Read the Guidelines for Using AI Tools in Academic Assignments for more details.

What is a prompt?

In generative artificial intelligence models, a prompt refers to the textual input supplied by users to direct the model’s output.

Basic prompts can be as simple as asking a direct question or providing instructions for a specific task. For example, with image generation models like DALL·E 3, prompts are often descriptive, while in large language models (LLMs) like ChatGPT-4 or Gemini, they can be simple queries. Advanced prompts involve more complex structures, such as "chain of thought" prompting, where the model is guided to follow a logical reasoning process to arrive at an answer.

Why do we need to write prompts?

Prompts are the primary way to communicate intentions, context, and goals to a language model. Unlike search engines or a library database that return existing documents, an AI generates responses based on your input. It is creating the text rather than retrieving content.

How does it do this? Through a process called tokenization.

Tokenization plays a foundational role in how artificial intelligence understands and processes human language. Imagine you're trying to teach a computer how to read. Before it can truly grasp meaning or respond intelligently, it first needs to break down the language into pieces it can manage. This is what tokenization does — it slices text into smaller units called "tokens" that the AI can analyze and work with.

These tokens are often words, but not always. Sometimes a token is a part of a word, like "un" in "unhappy", or even a punctuation mark. This breakdown allows the AI to process language more efficiently because it no longer has to consider an entire sentence as a single, overwhelming chunk. 

Here is an example of tokenization from the ChatGPT tokenizer.

By converting language into a sequence of tokens, AI models can learn patterns, relationships, and context based on how tokens commonly appear together. For instance, in the phrase "machine learning", the AI learns that "machine" is often followed by "learning", which helps it predict and generate coherent text.

Are responses to prompts always accurate?

All AI models are prone to making mistakes, usually referred to as hallucinations.

Tokenization and hallucinations in AI models like ChatGPT are subtly connected. Tokenization, the process of breaking text into smaller units, forms the basis for how the model understands and generates language. When this process mismatches the model’s training patterns, it can indirectly contribute to hallucinations—outputs that are fluent but factually wrong or entirely made up.

Hallucinations often stem from how the model interprets token patterns. Since models learn by predicting the next token, any noise, inconsistency, or bias in the training data can create misleading associations. For example, frequently co-occurring medical terms may cause the model to confidently generate incorrect diagnoses if those patterns aren’t well-grounded.

Tokenization also impacts the clarity of the model’s understanding. Models don’t understand words like humans do—they rely on token patterns. If tokenization fragments a word, especially a rare or technical one, the model might lose its meaning, resulting in distorted but plausible-sounding output.

The quote from this article sums up the problem;

[a Large Language Model] is a system for haphazardly stitching together sequences of linguistic forms it has observed in its vast training data, according to probabilistic information about how they combine, but without any reference to meaning (Bender et al., 2021, p. 617)

In short, while tokenization doesn’t directly cause hallucinations, it shapes how models process input and make predictions. Small issues can lead to significant errors. While clear, well constructed prompts can help reduce the possibility of hallucinations, it can't prevent them.

So remember, always verify any information that is generated by AI, regardless of how confident the response may sound!

Introduction

This guide is focused on creating prompts for publically available Large Language Models (LLM's) like ChatGPT, Claude, Gemini or LLaMA. These methods will not be effective for academic models like Scite, Elicit, Consensus or Scopus AI as they employ a different method, called Retrieval Augmented Generation (RAG), to return answers from sources such as web or database searches. It also focuses on text prompts rather than multimodal prompts involving images, audio or video.

Core Principles of Prompting

Understanding how to interact effectively with a LLM such as ChatGPT begins with two fundamental principles:

1. Non-Determinism
   Large Language Models generate responses probabilistically. This means that identical prompts given at different times may produce varying outputs. Variability is not necessarily a flaw, it reflects the generative flexibility of the model.

2. Iterative Nature of Prompting
   Prompting is an interactive and evolving process. A first attempt may not yield the ideal result. Learning to revise prompts, sometimes through several iterations, is a critical skill in obtaining accurate, useful responses from an AI model.

Zero-Shot Prompting

A basic prompt consists of a single, clear instruction or question but no examples are provided to illustrate the desired behaviour. This is known as zero-shot prompting.

With zero-shot prompting the goal is to provide a clear instruction to the AI using plain, specific language that communicates one task or question at a time. In general the model relies on its pre-existing knowledge acquired during training to interpret and carry out a zero-shot prompt, making it suitable for tasks like definitions, factual queries, or short explanations.

Key Characteristics of Basic Prompts

1. Clarity and Specificity
   Use precise language to avoid ambiguity. Define the task clearly using concrete terms.
   Example:

   • Vague: “Tell me about climate change.”

   • Better: “Define the term ‘climate change’.”

2. Single Task Focus
   Ask about one concept or objective at a time. This helps the model stay on topic and deliver concise results.
   Example: Ask for a list or definition, not both.

3. Minimal or No Context
   Basic prompts do not require background information. When a prompt is clear and well-formed, the AI can rely on its existing knowledge base.

Why Start with Basic Prompts?

Basic prompts form the foundation of prompt engineering and are ideal for:

• Rapid retrieval of facts or definitions

• Exploring how the AI interprets straightforward instructions

• Practicing prompt structure before moving to complex tasks like reasoning, dialogue, or formatting

They are also highly effective for tasks like topic brainstorming, quick summarization, and understanding technical terms.

Examples of Basic Prompting

Each example below illustrates the best practices for basic prompting and includes the rationale for its effectiveness.

A Simple Definition

Prompt: “What is the definition of plagiarism in academic writing?”
Why it works: It is a single, direct question targeting a known concept. The model can respond with a precise, focused definition without requiring extra guidance.

A List 

Prompt: “List three key benefits of study groups for college students.”
Why it works: The prompt specifies a quantity (“three”) and a topic (“study groups”). It sets clear expectations for the structure and scope of the response.

A Basic Explanation

Prompt: “Explain the principle of supply and demand in one paragraph.”
Why it works: This request introduces a mild constraint (paragraph format) while staying focused on a single economic principle. It allows the AI to generate a brief yet structured explanation.

Brainstorming

Prompt: “Give five research paper topics related to renewable energy.”
Why it works: This prompt invites idea generation but restricts it to five items on a specific subject. This balance between creativity and structure is ideal for entry-level brainstorming tasks.

Limitations and Opportunities

While basic prompts are a helpful starting point, they do have constraints:

• They may lack depth without context or follow-up questions.
• For complex tasks, you may need to shift to few-shot or chain-of-thought prompting, where the model is shown examples or guided through reasoning steps.

Nevertheless, mastering basic prompting lays the groundwork for more sophisticated techniques such as prompt chaining, style-based formatting, or iterative self-correction patterns.

Learning Through Simplicity

By focusing on clarity, you can gain insights into how AI systems interpret natural language.

As you become more comfortable with basic prompting, you will be better prepared to explore more advanced prompt strategies that allow for deeper, more nuanced interactions with AI tools.

Introduction

Intermediate prompts build upon the foundation of basic prompting by adding more context, structure, or complexity to the request. At this level, prompts become more detailed and may guide the AI to produce richer, more nuanced responses.

Characteristics of intermediate prompts

Intermediate strategies maintain clarity while handling multiple aspects or a more complex single task. In practice, this often means an intermediate prompt might do one or more of the following: set a specific persona for the AI, describe the intended audience or style of the response, impose a format or add data, or combine two related tasks in one prompt. 

Lets look at each of these in detail.

Personas or Roles

You can assign the AI a role to play, sometimes called a persona, to influence the style and content of the response. For example, a prompt might begin with “Imagine you are a tutor explaining...” or “You are a historian; analyze the following event...”. By giving the AI a persona, you can steer the response’s tone and depth.

You can try a fun example like the prompt below!

This technique enriches the interaction, as the AI will tailor its answer to fit the given context or persona. For instance, telling the AI “You are a textbook author for beginners” before asking for an explanation ensures the answer will likely be simpler and more pedagogical in tone than a generic explanation.

Structured and Multi-part Instructions

Unlike basic prompts that focus on one question, intermediate prompts may ask the AI to handle a sequence of related tasks or a question with multiple facets.

The prompt might explicitly structure what is needed – for example:

“Summarize the main point of the attached article on prompting, then give two reasons why it matters.”

Here, the AI must perform two steps in one response (summary + implications). Multi-part prompts remain clear but expect the model to produce a more organized answer covering each part.

This approach leverages the AI’s ability to follow a structured request, giving you a response that is logically organized according to the format.

Specific Formatting and Output Guidelines

You can specify the desired format or length of the output beyond what a basic prompt would do.

You can request a certain structure (bullet points, an outline, a paragraph, etc.) or indicate the audience reading level. For example:

“Make a basic outline for an essay about online privacy, using Roman numerals for the main sections and bullet points for subsections”

By stating the format (outline with specific style), the prompt guides the AI to produce a well-structured response. Similarly, one could say;

“Explain the theory in plain language as if addressing a high school student.”

Prompts like this ensure the content is pitched at the right level of complexity and clarity. These additional instructions lead to more tailored and useful responses.

Inclusion of Supporting Context

At the intermediate level, a prompt might include brief background information or data that the AI should use. For example:

“Given the following data attached below on city population growth, identify two trends and suggest possible causes.”

The prompt would then include a summary of the data or key figures. By embedding context, like a scenario or data snippet, you can help the AI deliver a response grounded in that context, rather than a generic answer. This approach is especially useful for tasks like analyzing a passage or interpreting data, where the prompt itself can supply the needed information for the AI to work with.

These strategies allow intermediate prompts to handle more complex tasks than basic prompts, without overwhelming the AI or sacrificing clarity. The prompts remain explicit about what is required, but they guide the AI towards a more detailed and relavent response.

When you add complexity, make sure the wording of the prompt stays unambiguous and logically ordered, perhaps by breaking the prompt into steps or bullet points if necessary.

Further Examples

Below are more examples of intermediate prompt and explanations:

Role-Based Explanation Prompt

Prompt: “You are a writing tutor. Explain the importance of a thesis statement in an essay, and give an example of a strong thesis.”
Explanation: This prompt assigns the AI a specific role, a writing tutor, which frames the style of the answer. It also has a two-part instruction: first explain the importance, then provide an example. The role implies the response should be didactic and supportive, as a human tutor would be. Because of the role and structured request, the AI’s answer will likely be more detailed and instructive than a basic explanation – it will explain why thesis statements matter and illustrate with an example. 

Audience-Specific Explanation Prompt

Prompt: “Explain the concept of opportunity cost in simple terms for a high school student, and provide a real-life example.”
Explanation: In this case, the prompt specifies an audience (a high school student) and asks for simplicity in the explanation. It also asks for an example to illustrate the concept. These additions elevate the prompt from a basic definition to an intermediate-level task that demands clarity (no jargon, simplified language), and completeness (the example reinforces the explanation). This prompt should produce an answer that is accessible and illustrative, likely starting with a plain-language definition of opportunity cost and then describing a relatable scenario as an example.

Structured Outline Prompt

Prompt: “Outline a research proposal on the effects of social media on academic performance. Include sections for Introduction, Literature Review, Methodology, and Expected Results.”
Explanation This prompt explicitly asks for an outline with specific section headings. It provides a clear structure that the AI should follow, turning what could be a broad task into a formatted plan. By naming the sections, it ensures the AI’s response will be organized into those parts. The AI is effectively guided to fill in a framework. Under each of the mentioned sections, it should give a brief description or key points relevant to the topic.

Comparative Analysis Prompt

Prompt: “Compare and contrast Theory X and Theory Y regarding their approaches to motivation. Organize your answer into two paragraphs: one for similarities and one for differences.”
Explanation: Here the prompt asks the AI to perform an analytical task – comparing two theories – and also specifies how to structure the answer (by similarities vs. differences, each in its own paragraph). This intermediate prompt ensures the AI covers multiple facets of the question methodically. The clarity about structure (“two paragraphs”) and content (what each paragraph should focus on) guides the reasoning process. The AI’s response will likely first discuss common points between Theory X and Y, then clearly delineate how they diverge, following the prompt’s requested format. This example illustrates how intermediate prompts can direct the AI to engage in higher-order reasoning while still providing a scaffold that keeps the response organized and on-topic.

A Prompt Canvas

Michael Hewing and Vincent Leinhos have developed a Prompt Canvas that provides a good outline of these and more intermediate methods.

Through these examples, we can see intermediate prompts, by giving context, structure, or multi-step instructions. allow us to obtain more comprehensive and tailored answers.

Advanced Prompting Strategies

Advanced prompts represent the most sophisticated level of interaction with generative AI. At this stage, prompts may become quite elaborate, incorporating techniques to handle complex tasks that require reasoning, adaptation, or multi-step processes.

Advanced prompt strategies often involve guiding the AI’s thought process or providing it with examples to learn from within the prompt itself. The focus is on leveraging the AI’s capabilities for in-depth analysis, creative problem-solving, and iterative refinement of answers.

You can read more about the methods below and others in these papers.

• Prompt Design And Engineering: Introduction And Advanced Methods
• An Empirical Categorization of Prompting Techniques for Large Language Models: A Practitioner's Guide
• Towards a Catalog of Prompt Patterns to Enhance the Discipline of Prompt Engineering
• The Prompt Report: A Systematic Survey of Prompting Techniques
• A Prompt Pattern Catalog to Enhance Prompt Engineering with ChatGPT

Few-Shot Prompting

One advanced pattern is few-shot prompting, where the prompt includes one or more examples of the task at hand to “teach” the model what is expected. Instead of relying purely on the AI’s pre-trained knowledge, as in zero-shot prompts, you supply demonstrations in the prompt itself.

For example, before asking the AI to answer a question or produce text in a certain format, the prompt can show a couple of Q&A pairs or formatted examples:

“Example: [Question] – [Answer]. … Now, given this new question, provide the answer in the same format.”

By providing two or three examples, called exemplars, the prompt helps the model recognize the pattern or style required, leading to more accurate and consistent outputs.

Here is an example:

Initial Prompt to the AI:

Q1: Lily has 3 packs of pencils. Each pack has 5 pencils. She gives 4 pencils to her friend. How many pencils does she have now?
A1: First, we find out how many pencils Lily starts with.
3 packs × 5 pencils per pack = 15 pencils.
Then she gives away 4 pencils.
15 - 4 = 11 pencils.
Answer: 11

Q2: A bakery made 24 cupcakes. They packed them into boxes with 6 cupcakes in each box. Then they sold 2 boxes. How many cupcakes are left?
A2: First, calculate how many boxes were packed.
24 ÷ 6 = 4 boxes.
They sold 2 boxes, which is 2 × 6 = 12 cupcakes.
24 - 12 = 12 cupcakes left.
Answer: 12

Q3: A farmer has 12 cows and 8 goats. He sells 5 cows. How many animals does he have now?
A3: Start with total animals:
12 cows + 8 goats = 20 animals.
He sells 5 cows.
20 - 5 = 15 animals.
Answer: 15

Now test the model with a new question:

Q4: Sarah has 10 apples. She buys 2 more bags of apples, each with 4 apples. Then she eats 3 apples. How many apples does she have left?

Few-shot prompts can also be used to guide the AI in following a specific writing style or analytical method. For instance, one could include a sample excerpt and its summary, then prompt the AI to summarize a new excerpt in a similar manner including the tone and style. This strategy leverages the model’s in-context learning ability. Essentially, the AI learns from the examples on the fly and applies this understanding to the new query.

Chain-of-Thought

A hallmark of advanced prompting is explicitly prompting the AI to perform multi-step reasoning or to show its work. Chain-of-thought (CoT) prompting is a technique where the prompt encourages the AI to break down a complex problem into a sequence of logical steps, rather than jumping straight to an answer. Research has shown that chain-of-thought prompting can significantly improve accuracy on multistep problems by preventing the AI from making leaps of logic.

A clever version of this method was developed, called Zero Shot CoT. All you have to do is add the following sentence at the beginning of a more complex or multi-step prompt and the AI will try to break down the prompt into steps without having an example or further instructions.

"Let's think step by step"

Self-Critique and Iterative Refinement

Advanced prompting can also involve asking the AI to critique or improve its own answers. After an initial response, you might try a follow-up prompt:

“Identify any errors or gaps in the above answer and correct them.”

Alternatively, the prompt itself can anticipate this by saying,

“Give your answer and then evaluate its accuracy.”

Or you can try a more complicated prompt that asks for a detailed analysis of previous answers.

For all of the previous responses, provide a bulleted list that includes (1) any assumptions you are making, (2) citations of any sources you have used to determine the response to my prompt, and (3) any biases or concerns you want to review if you had access to unlimited data or information for each response.

This is related to self-criticism prompting techniques, where the model is guided to reflect on its output and refine it. By prompting an LLM to double-check or justify its response, you can often obtain more reliable and well-reasoned results.

This iterative approach mirrors a high-level process – not just answering a question, but critically evaluating the answer. It harnesses the AI’s ability to perform internal checks ultimately improving the accuracy and quality of the final output.

Complex Prompt Structures and Chaining

At the advanced level, prompts can become quite complex, sometimes resembling a small script or program for the AI.

They may combine multiple advanced elements – for instance, a prompt might set a role, provide context data, include an example, request a step-by-step solution, and even ask for a final answer in a specific format all in one prompt. Crafting these prompts requires a deep understanding of how the AI interprets instructions. Additionally, you can employ prompt chaining, where instead of one giant prompt, they use a sequence of prompts to tackle a complex task in stages.

A complex example of this method can be seen in NASA's Bio-Inspired Design and Research Assistant (BIDARA), a prompt that extends over several pages!

For instance, you might first prompt the AI for an outline of a research paper, then on the next prompt ask for a draft of one section based on that outline, and so on. This sequential approach (planning    execution  refinement) is an advanced use pattern that can yield highly customized results.

Remember! Always Verify Responses

Given these advanced techniques, it’s important to remember that with greater prompt complexity comes the need for careful verification of AI outputs. Complex prompts can generate very convincing, detailed answers, but they might also propagate serious errors. Always cross-check facts and logic, adjusting your prompts and iterating to get better results.