Asking better questions: the art of LLM prompting

Getting the most out of Large Language Models (LLMs) isn’t just about asking questions, it’s about asking the right kind of questions. As your prompts evolve from basic queries to more advanced reasoning, so does the power of the LLM’s responses. 

In this article, we’ll break down how you can progress from simple prompts to more sophisticated techniques and explore why these methods work from a technical standpoint. 

Ask and you shall receive 

Simple queries are where most users begin. This type of prompting is ideal for fact-based questions where the answer is well-known and straightforward. Think of it as the AI equivalent of looking something up on Wikipedia. These queries are used in everyday applications like virtual assistants, customer service bots, or just to quickly retrieve a piece of information. 

For example, asking an LLM to retrieve a historical fact or define a common term is a perfect use case for simple queries. They’re fast, effective, and don’t require any advanced setup. 

Example: 

Prompt: “What’s the speed of light? 

Response: “299,792,458 meters per second” 

Why it Works: LLMs, especially models like GPT, are trained on vast datasets containing general knowledge. When you ask a factual question, the model’s attention mechanism quickly locates high-probability answers based on patterns it has encountered. This is a simple retrieval operation – no complex reasoning is needed. The LLM identifies the most statistically relevant token sequences and generates a response almost immediately. 

A little context goes a long way 

Once you’ve mastered simple queries, you’ll often want more from the model. Intermediate prompting involves providing additional context or setting constraints that shape the model’s response. This is useful for more specific questions, where you need to guide the LLM to focus on specific information or format the output in a particular way. 

This technique is often used in scenarios where the prompt requires a more tailored response – such as summarising an article, generating reports, or offering analysis based on specific parameters. Adding context helps narrow the focus, while setting constraints ensures the response is aligned with the desired output length or format. 

Contextual prompts 

By giving the model extra information, you guide it to focus on a particular area of interest. 

Example: 

Prompt: “In the context of surviving a zombie apocalypse, why is bringing a rubber duck essential? 

Constraining responses 

Limiting response length or format allows for better control over the output. 

Example: 

Prompt: “Explain how coffee is made, but in 10 words and like you’re a caveman.” 

Why it Works: LLMs rely heavily on self-attention mechanisms, allowing the model to focus on specific parts of the input context while generating the response. By giving it context, you’re helping the model “weigh” certain information more heavily.  

Constraints trigger the model’s token management system, where it limits or extends the output based on patterns learned from its dataset about brevity or elaboration. This is how the model knows to cut off at 10 words while still being coherent. 

Getting creative (or organised) 

Once you’re comfortable with shaping simple responses, you can use LLMs for creative tasks or to follow specific formats. This kind of prompting is where LLMs begin to show their full range – able to create original content or generate structured outputs based on a given format. 

Creative prompts are ideal for writing stories, generating ideas, or even crafting marketing copy, while structured prompts work well in tasks like creating reports, essays, or formal documents. In both cases, the LLM builds on its understanding of narrative, structure, and language patterns. 

Creative prompts 

Tapping into the model’s creativity allows for generative outputs like short stories or poetry. 

Example: 

Prompt: “Write a love letter from a toaster to a piece of bread” 

Structured prompts 

You can ask the LLM to follow specific formats, like essays or debates. 

Example: 

Prompt: “Write a 5-sentence movie review of Shrek as if it were a Shakespearean play.” 

Why it Works: LLMs are trained on diverse textual structures – novels, articles, essays – so they have an inherent understanding of different formats. For creative tasks, the model draws on latent representations of concepts learned during training, combining them in new ways to generate novel outputs. Structured prompts rely on the model’s ability to identify and follow common writing patterns, such as introductions and conclusions, based on sequence models that handle textual coherence. 

One step at a time 

Multi-step reasoning is essential when you want the LLM to handle more complex tasks that require logic, calculations, or breaking down a problem into smaller parts. This method is particularly useful in domains like programming, financial analysis, or solving puzzles where the answer isn’t immediately obvious. 

In multi-step reasoning, the model doesn’t just generate an answer – it explains the process step-by-step. This makes it perfect for scenarios where transparency is needed or where the complexity of the problem requires a detailed breakdown. 

Example: 

Prompt: “If I buy 12 donuts and eat 5, how many donuts are left for tomorrow? Explain each step.” 

Why it Works: Multi-step reasoning leverages the transformer architecture of LLMs, particularly the way it tracks dependencies across different tokens. The model follows your instructions step by step, maintaining context and order throughout the process. By instructing it to explain each stage, you engage the model’s ability to model logical flows and generate responses that reflect those dependencies across a sequence of operations.

Show, Don’t Tell 

Few-shot prompting is a powerful technique where you provide the model with a few examples to help it understand the task before it completes the next one. This method is useful in tasks where you need the LLM to mimic a particular style, pattern, or behaviour. Few-shot learning is often employed in translation, summarisation, or specialised content generation. 

This approach is particularly useful when the task requires the LLM to perform in a way it may not have been explicitly trained for – offering the model a handful of examples can dramatically improve accuracy and consistency. 

Example: 

Prompt: “Translate these sentences: 

The cat is on the sofa. 

The dog is chasing its tail. 

Now translate: The parrot is plotting world domination.” 

Why it Works: Few-shot learning taps into the LLM’s ability to generalise from examples. The model uses the provided examples to infer patterns and apply them to unseen tasks. This is an application of transfer learning, where the model applies previously learned information (in this case, translation rules) to new queries. The model is already primed for flexibility, and the few examples help it narrow down the context and style needed for the task. 

Let’s Think This Through 

Chain-of-Thought (CoT) prompting is the ultimate technique for engaging the model in deep, multi-step reasoning tasks. In CoT, the LLM doesn’t just provide an answer; it walks through each step of the process, explaining its reasoning along the way. This is essential for tasks involving logic, mathematical calculations, or decision-making, where each part of the process must be considered in sequence. 

CoT prompting is particularly useful in fields like programming, legal reasoning, or any situation where clarity and logic are crucial. By forcing the model to articulate its thought process, CoT ensures greater accuracy in complex problem-solving. 

Example: 

Prompt: “If you’re 30 years old and your brother is half your age, how old will your brother be when you’re 60? Walk through each step.” 

Why it Works: CoT prompting exploits the model’s self-attention layers, ensuring that it doesn’t skip steps in reasoning. The key to CoT’s success lies in its ability to handle intermediate reasoning states, where the model must process each part of the problem independently before connecting them into a coherent whole. This technique significantly improves performance on tasks that require logical deductions or sequential reasoning, reducing errors by forcing the model to focus on each stage of the problem before arriving at a conclusion. 

The Evolution of Effective Prompting 

As you’ve seen, prompting is not a one-size-fits-all approach. Each level – whether basic, intermediate, or advanced – draws on different strengths of the LLM’s underlying architecture. By understanding how these models handle context, structure, and reasoning, you can craft prompts that maximise the AI’s capabilities, from simple questions to complex problem-solving.