Mastering the Art of Prompting

February 21, 2019

Mastering the Art of Prompting: A Deep Dive into Chain-of-Thought and Other Techniques

The advent of Large Language Models (LLMs) like GPT-3 and ChatGPT has opened a plethora of possibilities, from drafting emails to coding software. But despite their potential, LLMs aren't psychic; they require skillfully crafted prompts to deliver meaningful outputs. This blog post explores the fascinating world of prompting techniques, spotlighting the intriguing "Chain of Thought" approach.

What is Prompting?

Before diving into the complexities, let's start with the basics. A "prompt" is essentially the input you give to an LLM. It guides the model in generating an output that ideally matches your expectations. However, the way you craft this input can significantly affect the model’s performance.

Few-Shot Learning

One of the most straightforward methods of prompting is "few-shot learning." Here, you provide a series of examples before your main question, helping the model grasp the context. For instance:

plaintextCopy code

Example 1: Translate "hello" into French.
Answer 1: bonjour

Example 2: Translate "thank you" into French.
Answer 2: merci

Translate "goodbye" into French.

Zero-Shot Learning

In zero-shot learning, you don't provide any examples, just the task you want the model to perform. This method relies on the model's pre-trained knowledge to interpret the prompt and respond appropriately.

Chain-of-Thought Prompting

Now, onto the showstopper: Chain-of-Thought (CoT) Prompting. Unlike standard prompts that aim for a direct answer, CoT prompts guide the model through a sequence of logical steps to arrive at a conclusion. This is particularly useful for tasks requiring critical reasoning or complex problem-solving.

Variants of CoT Prompting

  1. Few-Shot CoT: Similar to few-shot learning, this method provides a sequence of examples showing the step-by-step reasoning before asking the model to perform a similar task.
  2. Zero-Shot CoT: Utilizes a trigger phrase that instructs the model to display its reasoning steps without any prior examples.
  3. Faithful CoT: A more advanced form, this includes multiple steps and toolsets to ensure the model's output accurately reflects its reasoning pathway.

Why Chain-of-Thought?

  1. Transparency: CoT prompts make it easier to understand how the model arrived at a particular conclusion.
  2. Error Mitigation: By breaking down the reasoning process, it becomes simpler to identify and correct errors.
  3. Enhanced Reliability: Knowing the logical steps involved adds an extra layer of trust in the model's outputs.


Prompting is an art form that can unlock the true potential of LLMs. While few-shot and zero-shot learning are invaluable tools in your prompting toolkit, Chain-of-Thought prompting offers a unique avenue for tasks that require a greater depth of reasoning. As LLMs continue to evolve, mastering these techniques will become increasingly vital in leveraging their full capabilities.

AI can solve your organisations biggest problems.
We don't charge for our advice, so drop us a line
Start Now