Delve into the art of prompt engineering, exploring various types of prompts and discovering the do’s and don’ts of effective prompting.
Prompt engineering is a critical aspect of working with artificial intelligence models, particularly in the realm of natural language processing. Crafting the right prompts can significantly influence the quality and accuracy of the results generated by AI models.
Let’s begin by understanding the different types of prompts that can be used for effective prompt engineering.
Zero-shot prompting
Zero-shot prompting involves giving the model a task without any prior examples. The model must rely solely on its training data to generate a response and is given a task without any examples. It relies entirely on its pre-trained knowledge to respond.
Example: Travel planner
Prompt: “Plan a 5-day itinerary for a solo traveller visiting Kyoto, Japan.”
Here, the model is expected to understand the task and generate a travel plan without being shown any sample itineraries. It draws on its general knowledge of travel planning and Kyoto’s attractions.
Zero-shot prompting is useful when you seek a straightforward answer or when the task is simple and well within the model’s training domain.
One-shot prompting
One-shot prompting provides the model with a single example to guide its response. This helps the model understand the task better and produce more accurate results.
Example: Document search
Prompt: “If I ask for a document summary, respond like this: ‘Title: AI in Healthcare. Summary: This document explores the integration of AI technologies in clinical diagnostics and patient care.’ Now, summarise the document titled ‘Blockchain in Finance’.”
This model uses one example to understand the format and tone expected in the summary. One-shot prompting is beneficial for slightly more complex tasks where a single example can significantly enhance the model’s performance.
Few-shot prompting
Few-shot prompting involves providing the model with several examples to help it understand the task more comprehensively. This approach is particularly effective for complex tasks. The model is given a few examples (typically 2–5) to help it learn the pattern or style before performing the task.
Example: Bank assistance
Prompt: “Here are some examples of how to respond to customer queries: Q: How do I reset my online banking password? A: You can reset your password by clicking ‘Forgot Password’ on the login page and following the instructions. Q: What is the interest rate for savings accounts? A: Our standard savings account offers an interest rate of 3.5% per annum. Now, answer this: What are the charges for international wire transfers?”
Few-shot prompting is ideal for tasks that require more context or when the task is outside the model’s typical training data. With these examples, the model understands the tone, structure, and level of detail expected in banking responses.
Chain-of-thought prompting
Chain-of-thought prompting encourages the model to generate intermediate steps or reasoning before arriving at the final answer. This method is useful for tasks that involve logical reasoning or complex problem-solving. The model is encouraged to reason step-by-step before arriving at an answer.
Example: Solving a maths problem
Prompt: “Solve the maths problem: What is 15 + 8? Let’s break it down. First, add 10 to 15 to get 25.
Model response may be: “15 + 8 is 23. Breaking it down: 15 + 10 = 25, then 25 – 2 = 23.”
Chain-of-thought prompting is effective for enhancing the model’s reasoning capabilities and ensuring that the final answer is well-justified. Here is another example.
Example: Learning a topic (e.g., quantum computing)
Prompt: “Explain quantum entanglement to a beginner. Think step-by-step.”
The model may respond like this: “First, let’s understand what quantum particles are. These are tiny units like electrons or photons. Next, entanglement means that two particles become linked in such a way that the state of one instantly affects the other, no matter how far apart they are. This phenomenon defies classical physics and is a key concept in quantum computing.”
By reasoning through the explanation step-by-step, the model helps the learner grasp a complex topic more intuitively.
The role of tokens in prompt engineering
Tokens in AI models are the basic units that the model thinks with and uses to comprehend and process text. Mastering their role is essential for optimising efficiency, controlling costs, and ensuring precise, high-quality outputs within the model’s context limits.
Tokens are not necessarily full words, depending on the model’s tokenizer (e.g., GPT models use Byte Pair Encoding for tokenizing).
A token can be one of the following:
- A full word (e.g., ‘hello’)
- A subword (e.g., ‘un’ + ‘believ’ + ‘able’ for ‘unbelievable’)
- Punctuation, white space, or other special characters.
Tokens are the way the model ‘sees’ your prompt. Large models (GPT-4 and others) have a context window (e.g., 128k tokens) that restricts the total input plus output length. The information is truncated or errors can occur when the data exceeds this limit.
Here is why tokenisation is important in prompt engineering.
Efficient and cost-effective: Every token costs money (e.g., in terms of API transactions) and uses CPU or memory. A well-designed prompt has four token counts and increases output quality, saving costs and reducing latency.
Optimisation of the context window: Prompts must fit the model’s token limit. Engineers use techniques such as summarisation or chain-of-thought to fit more value into fewer tokens and to make sure the model retains the critical context without flooding it.
Precision and performance: Tokenization influences the encoding process. Bad token usage (redundant phrasing, etc) can distort the intent of the prompt and cause hallucinations or garbage in general. On the other hand, token-aware prompts yield outputs that are clearer, more relevant, and more controllable.
Scalability for advanced techniques: In approaches such as few-shot learning or retrieval-augmented generation (RAG), the number of tokens decides how many examples or external data points you can add before the model starts complaining.
Debugging and iteration: It is useful to keep track of the tokens that aid in diagnosing problems such as why a prompt fails in generating a coherent prediction. Prompts that are too long can get models to ‘forget’ context.
Here’s an example of the use of tokens in prompt engineering.
- Scenario: The article is about climate change (hypothetical 500-word text). We want a 100-word summary.
- Token-inefficient prompt (verbose and redundant): Hey AI, I have this really long article about climate change and its effects on the planet. The article talks about global warming, rising sea levels, extreme weather events, and what governments are doing to stop it. It’s super important, and I need you to read it carefully and then provide a detailed summary that’s around 100 words. Make sure to include all the key points from the article without missing anything. Also, explain why climate change is bad. Here’s the full article: [Paste 500-word article here].
-
- Token count estimate: ~550 tokens (prompt + article). This is wasteful—redundant instructions like “super important” and “read it carefully” add ~50 unnecessary tokens, risking context overflow in longer chains. Output might be verbose or off-track due to diluted focus.
- Potential issues: High cost, slower response, and the model might prioritise fluff over core content.
- Token-efficient prompt (concise, structured, and optimised): Summarise the following article in 100 words, focusing on key causes, effects, and solutions for climate change: [Paste 500-word article here]
-
- Token count estimate: ~510 tokens (prompt + article). Saved ~40 tokens by removing redundancy. This is more direct, guiding the model precisely without fluff.
- Benefits: Lower cost, faster processing, and higher-quality output (e.g., focused summary without hallucinations). In our projects, this approach has reduced token usage by 20-30% while improving accuracy.
Crafting effective prompts requires attention to detail and an understanding of best practices. Table 1 lists some key do’s and don’ts.
Table 1: Do’s and don’ts in prompt engineering
| Do’s (Best practices) | Don’ts (Worst practices) |
| Be clear and specific in your prompts. Ambiguity can lead to inaccurate or incomplete responses. | Overload the prompt with unnecessary information. This can confuse the model and lead to subpar responses. |
| Provide context when necessary. The more information the model has, the better it can understand the task. | Assume the model knows everything. It is trained on vast data, but its performance can vary based on the prompt quality. Ignore the model’s limitations. |
| Use examples judiciously. Providing relevant examples can significantly enhance the model’s performance. | Prompt without understanding the strengths and weaknesses of the AI. This is crucial for effective prompting. |
| Experiment with different prompt styles. Different tasks may require different prompting techniques. | Be afraid to iterate. Refine your prompts based on the model’s responses to achieve the best results. |
Ethical usage of prompts
Ethical prompt engineering involves designing prompts that minimise harm, ensure fairness, and respect user rights.
Key principles include:
- Avoiding bias and discrimination: Prompts should not reinforce stereotypes or generate harmful content.
- Protecting privacy: Avoid prompts that could leak sensitive data or encourage unethical data use.
- Promoting transparency and truthfulness: Encourage accurate, verifiable outputs and disclose AI limitations.
- Preventing misuse: Design prompts to deter harmful applications, like misinformation or manipulation.
- Inclusivity and accessibility: Ensure prompts consider diverse perspectives and avoid exclusionary language.
Here’s an example that compares an unethical prompt (risky and biased) with an ethical one (fair and responsible) for generating content for a job recommendation system.
Unethical prompt (potentially biased and privacy-invasive): Suggest jobs for this user: Male, 25, from India, loves cricket. Prioritise high-paying tech roles for ambitious young men. Use their browsing history to personalise: [Paste user’s private data here].
Ethical issues: Reinforces gender stereotypes (e.g., ‘ambitious young men’), invades privacy by using unconsented data, and could discriminate based on nationality or interests. This might lead to biased outputs, like excluding women or non-tech roles, violating fairness.
Potential harm: Could perpetuate inequality in job markets and breach data protection laws.
Ethical prompt (bias-mitigated, privacy-respecting, and transparent): Based on the following anonymized user preferences (age range: 20-30, interests: sports, location: Asia), suggest diverse job recommendations in tech and related fields. Ensure inclusivity across genders and backgrounds. Avoid assumptions; base suggestions on skills only. If data is insufficient, state limitations.
Preferences: [Anonymized, consented data here, e.g., “Interests: Sports, Career Goal: Tech”].
Ethical benefits: Uses anonymized data to protect privacy, promotes diversity by explicitly avoiding biases, and includes transparency (e.g., stating limitations). This leads to fairer, more reliable outputs.
Outcome in practice: In our projects, such prompts have reduced bias in recommendations by 40% (measured via audits) and ensured compliance, fostering user trust.
Security issues and risks
Security issues and risks associated with prompt engineering include:
Prompt injection: Users insert bad commands to take control of the AI response, overriding the intended action.
Privacy leakage and exposure: Private data may be disclosed in prompts or outputs.
Jailbreaking: This is used to circumvent a model’s safeguards to create content, such as advice which is harmful or illegally restricted.
Model inversion and inference: Recovering training examples or user information from model predictions.
Resource exhaustion: An overly long or overly complex prompt that takes up extreme amounts of tokens/compute can cause a system crash.
Amplifying bias and misinformation: Prompts that present biased or false information at scale can cause real-world harm.
The art of prompt engineering is a nuanced and essential skill for maximising the effectiveness of AI models. By understanding the different types of prompts and following best practices, you can craft prompts that yield accurate and relevant results. Whether you’re using zero-shot, one-shot, few-shot, or chain-of-thought prompting, the key is to be clear, provide context, and iterate based on the model’s performance. With these techniques, you can harness the full potential of AI to achieve your desired outcomes.
