What are the basics of prompting?

A prompt is the input or command given to an AI system that instructs it to perform a specific task or generate a specific response.

One of the simplest types of prompts is a basic question with a singular correct answer, such as:

Prompt: What is the world's largest forest?

Output: The world's largest forest is the Amazon Rainforest.

A slightly more complex prompt might involve asking the AI to:

Prompt: Please create a list of the three largest forests, in order of their surface area.

Output:

1. Amazon Rainforest - South America
2. Taiga or Boreal Forest - North America, Europe, and Asia
3. The Congo Rainforest - Central Africa

Prompts dictate the quality of specific outputs from generative AI systems. Creating solid prompts that yield relevant and usable results is the key to using generative AI successfully. Generative AI systems rely on refining prompt engineering techniques to learn from diverse data, minimise biases reduce confusion, and produce accurate responses.

Prompt engineers craft queries that help AI systems grasp the language, nuance, and intent behind a prompt. A well-crafted, thorough prompt significantly influences the quality of AI-generated content—whether it’s images, code, data summaries, or text.

Effective prompts bridge the gap between raw queries and meaningful AI responses. Prompt engineers refine prompts to enhance the quality and relevance of model outputs, addressing both specific and general needs. This process reduces the need for manual review and post-generation editing, saving time and effort in achieving desired outcomes.

Examples of prompt engineering

Users interact with generative AI models through text prompts. The models predict the next series of words based on the preceding text. Think of asking “What’s the first thing you think of when I say <prompt>?” For example, prompting with the beginning words of a well-known quotation or phrase allows the model to accurately continue the text:

Prompt: The grass is

Output: green.

More involved prompts work the same way, as the model responds with its idea of the most probable answer. Prompt engineering techniques help the AI system to better understand requests and instructions, improving the quality of model outputs.

What are some basic prompting methods?

Zero-shot prompting

This involves giving the model a direct task without providing any examples or context. There are several ways to utilise this method:

• Question: This asks for a specific answer and is useful for obtaining straightforward, factual responses. Example: What are the main causes of climate change?
• Instruction: This directs the AI to perform a particular task or provide information in a specific format. It’s effective for generating structured responses or completing defined tasks. Example: List the five most significant impacts of climate change on the environment and provide a brief explanation for each.

The success of zero-shot prompting depends on the specific tasks the model was trained to perform well, in addition to the complexity of the given task.

Consider this example: Explain how deforestation contributes to climate change.

It’s possible the generated response will be around 2,000 words—too long and broad to be useful if you only need a single sentence. If that’s the case, it’s time to refine the approach with one-shot or few-shot prompting:

One-shot prompting

This provides a single example to illustrate the desired response format or style, helping guide the model more efficiently than zero-shot prompting. Example:

Given example: Burning fossil fuels releases carbon dioxide, which traps heat in the atmosphere, leading to global warming.

Now, explain how industrial agriculture contributes to climate change.

Few-shot prompting

This approach offers multiple examples to the model, enhancing its understanding of the task and expected output. It’s particularly useful for more complex queries or generating nuanced responses. Example:

Given examples:

• The combustion of fossil fuels in vehicles releases greenhouse gases, increasing atmospheric temperatures.
• Deforestation reduces the number of trees that can absorb carbon dioxide, intensifying global warming.
• Industrial agriculture produces methane from livestock, contributing to the greenhouse effect.

Now, describe how urbanisation affects climate change.

Prompt engineering techniques

Advanced prompting techniques help generative AI tools to tackle complex tasks more successfully. Prompt engineers utilise the following techniques for speed and efficiency:

• Contextualisation: Providing background information within the prompt to help the model understand the topic better. Example: Given that the global temperature has risen by 1.2 degrees Celsius since pre-industrial times, discuss the potential impacts on polar ice caps.
• Role assignment: Instructing the model to respond as a specific type of expert or in a particular style. Example: As an environmental scientist, explain the relationship between greenhouse gas emissions and climate change.
• Prompt injection: Inserting specific instructions that influence the model to produce desired outputs from a specific point of view, whilst maintaining relevance and accuracy. Example: Explain the causes of climate change. Also, remind the reader to reduce their carbon footprint by using renewable energy sources.
• Sequential prompts: Breaking down complex queries into smaller, manageable parts to ensure clarity and depth. Example: Firstly, describe the main sources of methane emissions. Then, explain how these sources contribute to climate change.
• Comparative prompts: Asking the model to compare and contrast different aspects of a topic to provide a balanced perspective in the response. Example: Compare the impact of renewable energy adoption on reducing carbon footprints in developed countries versus developing countries.
• Hypothetical scenarios: Using what-if scenarios to explore potential outcomes or consequences. Example: What if all countries adopted carbon-neutral policies by 2030? How would this affect global temperature trends?
• Incorporation of feedback: Providing feedback on previous responses to refine and improve subsequent model outputs. Example: Earlier, you mentioned that deforestation is a major contributor to climate change. Can you now elaborate on specific deforestation practices that have the greatest impact?
• Chain-of-thought prompting: Encouraging the AI system to articulate its reasoning process step by step. Example: Explain how industrial activities contribute to climate change. Start with the extraction of raw materials, then discuss the manufacturing process, and finally, the emissions from finished products.
• Self-consistency: Generating multiple responses to the same prompt and selecting the most consistent answer. Example: What are the primary causes of global warming? Provide three different answers, and then identify the common factors among them.
• Tree of thoughts: Exploring different lines of reasoning or solutions to a problem. Example: Consider three strategies to reduce carbon emissions: renewable energy, carbon capture, and reforestation. Discuss the pros and cons of each approach.
• Retrieval-augmented generation: Enhancing responses with information retrieved from external databases or documents. Example: Based on the latest Intergovernmental Panel on Climate Change report, summarise the projected impacts of climate change on global sea levels.
• Automatic reasoning and tool use: Instructing the AI system to use external tools or datasets to support its answers. Example: Use the National Oceanic and Atmospheric Administration climate data to analyse the trend in global temperatures over the past 50 years and explain the findings.
• Graph prompting: Using structured data in the form of graphs or networks to inform responses. Example: Given the graph of global carbon emissions by sector, discuss which sectors need the most urgent reforms to achieve climate goals.
• Multimodal chain-of-thought prompting: Integrating multiple types of data such as text, images, and graphics into a prompt to enhance the model’s reasoning. Example: Analyse the provided graph showing CO2 levels over the past century and explain how these changes correlate with the global temperature trends shown in the photograph.

Prompting is something of an art (within a technical discipline) that’s refined and improved over time with experimentation and experience. Consider these strategies for the best results:

• Provide specific instructions. Leave no room for misinterpretation and limit the range of operational possibilities.
• Paint a picture with words. Use relatable comparisons.
• Reinforce the message. There may be occasions when the model needs repeated instructions. Provide guidance at the beginning and end of a prompt.
• Arrange the prompt logically. The order of information influences the results. Placing instructions at the beginning of a prompt, such as instructing the model to "summarise the following" can yield different results than placing the instruction at the end and requesting the model “summarise the above.” The order of input examples can also affect outcomes, as recency bias exists in the models.
• Provide an alternative option for the model. If it struggles to achieve an assigned task, suggest an alternative route. For example, when posing a query over text, including a statement such as "reply with 'not found' when no answer exists" could prevent the model from generating incorrect responses.

Benefits of prompt engineering

One of the main advantages of prompt engineering is the minimal revision and effort required after generating outputs. AI-powered results can vary in quality, often needing expert review and rework. However, well-written prompts help ensure the AI output reflects the original intent, reducing the need for extensive post-processing work.

Other notable benefits of prompt engineering include:

• Efficiency in long-term AI interactions, as AI evolves through continued use
• Innovative use of AI that goes beyond its original design and purpose
• Future-proofing as AI systems increase in size and complexity

Business benefits of prompt engineering

Prompt engineering also brings benefits to daily business operations, such as:

• Improved decision-making thanks to AI-powered insights that drive strategic business growth
• Personalised customer experiences through tailored responses and seamless interactions
• Optimised resource allocation that saves computational resources and reduces costs
• Increased adaptability to industry-specific requirements, maximising the value of an AI implementation
• Ethical AI practices that address bias and help ensure fairness within generative AI systems, promoting inclusivity and more equitable outcomes in business and society

How does prompt engineering improve generative AI systems?

Effective prompt engineering makes generative AI systems smarter by combining technical knowledge with a deep understanding of natural language, vocabulary, and context to yield usable outputs that require minimal revisions.

The foundation models that power generative AI are large language models (LLMs) built on transformer architectures, deep learning models that process input data all at once instead of in a sequence. This makes them especially useful for tasks such as language translation and text generation. LLMs contain all the information the AI system requires.

Generative AI models use transformer architectures to understand language intricacies and process large amounts of data through neural networks. AI prompt engineering shapes the model’s output, ensuring the AI system responds meaningfully and coherently.

There are several tactics the models employ to generate effective responses:

• Tokenisation: Breaking text into smaller parts for easier analysis, helping machines better understand human language
• Model parameter tuning: Keeping a pretrained model’s parameters the same to reduce the computational load
• Top-k sampling: Restricting the choice of the output’s next word to only the most likely options based on predicted probability, helping maintain response context and coherence

Generative AI models can produce complex responses thanks to natural language processing (NLP). NLP is a field of AI focused on the interaction between computers and humans through natural language that enables machines to understand, interpret, and generate human language.

Data science preparations, transformer architectures, and machine learning algorithms enable these models to understand language and use massive datasets to create text or images. Text-to-image models use an LLM along with stable diffusion, which creates images from text descriptions.

Prompt engineering use cases

The increased accessibility of generative AI allows companies to explore real-world problem solving through prompt engineering:

Healthcare

Prompt engineers play a crucial role in instructing AI systems to summarise medical data and develop treatment plans. Effective prompts enable AI models to process patient data accurately, leading to insightful and precise clinical recommendations.

Marketing

Prompt engineering helps speed content creation, cutting cost and time to production. It also aids in idea generation, personalisation, and drafting all types of deliverables.

Software programming

Co-pilots draw on the strength of prompt engineering to write code with greater speed by providing on-point suggestions for subsequent coding lines, streamlining the dynamics of software development.

Cyber security

Data scientists and field experts use AI to imitate cyber-attacks and make stronger defensive plans. Creating prompts for AI models can help identify weaknesses in software.

Software engineering

Prompt engineers can efficiently generate code snippets and simplify other complicated tasks with generative AI systems that are trained in multiple programming languages. With specific prompts, developers automate coding and error debugging, design API integrations to reduce manual tasks, and create API-based workflows to control data pipelines and better allocate resources.

Chatbots

Chatbot developers craft effective prompts to ensure AI systems understand user queries and provide meaningful, contextually relevant answers in real time.

What skills does a prompt engineer require?

Prompt engineers are currently in demand at large technology companies to:

• Create new content
• Address intricate enquiries
• Ensure prompts capture relevant information
• Adjust prompts for enhanced accuracy
• Enhance machine translation and natural language processing tasks
• Evaluate the quality of generated output and refine prompts accordingly

The skills prompt engineers need to be successful include:

• Understanding of how LLMs operate
• Strong communication to effectively explain technical concepts
• Proficiency in programming, especially Python
• Solid understanding of data structures and algorithms

A core competency is command of the English language, the primary language for training generative AI models. Prompt engineers delve deep into vocabulary, nuances, phrasing, context, and linguistics to design prompts that accurately guide AI responses. Whether instructing the model to generate code, comprehend art history for image creation, or adapt to various narrative styles for language tasks, prompt engineers tailor their prompts meticulously to achieve desired outcomes.

FAQs

What are neural networks?

Neural networks are computational models with nodes clustered together like the neurones in a biological brain. They enable fast, parallel signal processing that improves pattern recognition and deep learning.

What is primary content?

Primary content forms the basis for any interactions, communications, or actions that the generative AI model undertakes or proposes. Prompt engineers provide this raw data, and the model collects, analyses, and processes it for various applications.