A prompt is an instruction that you give a generative AI tool, which then produces an output (completion). Writing effective prompts (sometimes called prompt engineering) is a skill that requires practice (and a few strategies) in order for you to get the best results from the tool.
Below is an example of a prompt and output using ChatGPT 3.5:
The acronym ACCENT outlines six strategies to make your prompts more effective.
This framework for improving prompt engineering was adapted from the 26 guiding principles proposed by Sondos, M. B., Myrzakhan, A., & Shen, Z. (2024):
Alignment: Ensuring the model’s capabilities are in sync with your task.
Check if there another tool that matches your purpose more closely than the well-known large language models.
Your favourite software and platforms (for example, Microsoft Word) are now increasingly adding in AI tools to their functionalities (such as word or sentence completion). However, for some tasks, you might be better using specialised AI tools (e.g. for coding, generating images or searching for academic sources).
Clarity: Writing clear and concise prompts.
Tip: Make your prompt specific but not too long. Another benefit of being clear is that it can make you think through exactly what kind of output you want),
Context: Providing relevant context so that the output content is relevant to your needs.
Tip: Include relevant background information, including constraints (limiters).
Examples: Including examples of good or bad responses to help explain what you are looking for. This is sometimes called fewshot prompting.
Including examples to demonstrate the format of response that you require, or the action you want performed. This is also called multi-shot prompting.
Tip: you can upload a document containing the format you wish to use.
Further reading:
Neutral language: Design your prompts to minimise biases (for example, to minimise assumptions about gender or race).
Tip: Consider whether you need to use she/her or he/him. Often they/them will be sufficient.
Trial and error: Refining prompts based on the tool's performance and feedback.
Output from the tool can be extremely useful to pinpoint where your prompt is unclear, biased or could benefit from examples.
Output can also show where the tool has weak spots (e.g. providing inaccurate advice, or unable to perform the task).
Rubber duck prompting
Some people say that talking to a rubber duck (which won't ever answer!) can help you find a solution to your own problem simply by talking it through.
In the same way, asking questions (prompts) to an AI tool can help you to think through a problem for yourself. The benefit is not necessarily in the answers it gives you, but how the questions that you ask help you clarify your own thinking.
Further reading:
Asking AI to improve your prompts
Instead of using generative AI tools as search engines (e.g. asking them to provide factual information), try asking them to improve your prompts.
Other uses include asking for feedback on your writing (provide counterarguments, make your writing more concise).
Further reading:
Prompt chaining
Prompt chaining involves breaking down a task into steps and prompting the AI tool to complete each step, which is used in the following step. Prompt chaining "helps to boost the transparency of your LLM application, increases controllability, and reliability. This means that you can debug problems with model responses much more easily and analyze and improve performance in the different stages that need improvement." (DAIRAI)
Exceptions: This technique will not work when:
To understand what is possible with your tool, read the tool information carefully or run a few simple tasks before you begin a major task.
Further reading: