AI Literacy Toolbox: Prompt engineering 101

Prompt engineering is the art of crafting effective prompts to get the most useful and relevant responses from AI Large Language Models (LLMs). It involves carefully structuring and phrasing your requests to the AI in a way that gets the desired output.

Prompt engineering is particularly useful for AI research tools used for higher level research. See our AI for Research page for a comprehensive guide to these tools.

Quick links

Bias 

• It's important to be careful with the way you phrase your prompts to AI systems. Steer clear of prompt language that reinforces bias, perpetuates stereotypes or spreads false information. AI can amplify this bias depending on its training data and algorithms it was built on.
• In a tertiary setting one of your ethical obligations is to avoid bias in your language and to verify information. 

Privacy

• Don’t put personal information into generative AI (GenAI) that you wouldn't feel happy sharing with the world. This includes your own personal information, personal information relating to others, private business information or research which could be subject to a patent. Under current laws Tech companies will own this information.  

• If available, use privacy settings and forbid third parties from sharing any information you give GenAI.

• Do not use it to make important personal or business decisions, a human overview is essential. 

• Be critical and investigate any information GenAI gives you. By using its suggestions, you may be breaching privacy or copywrite laws, exposing yourself or your organisation to risk. 

• If you share GenAI-generated content with colleagues or third parties, you need to disclose the fact that you have used GenAI and be aware of its limitations prior to using it. Keep copies of your original documents, have a way of removing AI from documentation if you need to. 

• The legal landscape governing the use of GenAI is subject to change, so stay up to date with the rules of your institution and those of Aotearoa. 

Crafting effective prompts is a skill that involves understanding your subject matter, being aware of nuance, and knowing the strengths and limitations of the AI you're using. Just like choosing good keywords, the quality and specificity of your prompts directly impacts the relevance and accuracy of the AI's output.

One way to ensure high-quality output is to use The CLEAR framework. Designed by Leo S. Lo, it aims to make research interactions with AI models efficient. 

CONCISE

• Vague prompts will get vague responses: LLMs work on predictions using statistical likelihoods of the next word or token, a vague prompt can initially lead you off track multiplying inaccuracies. Use keywords (the main concepts of your question/topic) and avoid using unnecessary words.

LOGICAL

• If you have more than one concept, put them in a logical order: "Describe the elements of essay construction beginning with the introduction, then the middle paragraphs, and ending with the conclusion".
• If you have a complex task, break it down into a series of simpler questions (this effectively means you will give several follow up prompts): Just like keyword searching, breaking prompts up into sub tasks or steps, rather than searching a broad topic or typing in the question, will lead to better results. The more accurate and specific the better. Instead of prompting “revise this text”, be explicit, define the objectives, where the emphasis should lie, and define the constraints. Overly complex prompts may confuse the model.

EXPLICIT

• Be specific. Use keywords and instructions to guide the model and state the purpose of your prompt: Your query needs specific details and articulated outcomes. LLMs are refined by human-trained or reinforcement learning. Instead of asking for one suggestion, ask for three or five examples/analogies, and limit them to a paragraph or state a word limit for the answer. 

ADAPTIVE

• Adjust, fine tune, and refine as you go along: if you're not getting the responses you want, tweak your prompt. You may need to give it stronger parameters. Remember, specifics work best.

REFLECTIVE

• Is the output accurate? Is it relevant? Is it complete? Depending on the model you are using, accuracy and reliability can be an issue. You may need to verify the information.
• Some models are better than others. Experiment!

Source: Lo, L. S. (2023). The CLEAR path: A framework for enhancing information literacy through prompt engineering. Journal of Academic Librarianship, 49(4), Article e102720. https://doi.org/10.1016/j.acalib.2023.102720

A prompt comprises of two parts:

Part one

• The system prompt which establishes the context and goal;

•  and the role AI and the user will play.

Prompt:

Part two

• Lists the specific requirements of the answer including scope and boundaries, and greater detail. 

Prompt

Why this is important

LLMs work on predictions using statistical likelihoods of the next word or token. A vague prompt can initially lead you off track and multiply inaccuracies. 

Just like keyword searching breaking prompts up into sub tasks or steps rather than searching a broad topic or typing in the question will lead to better results. The more accurate and specific the better. Instead of prompting “revise this text”, be explicit, define the objectives, where the emphasis should lie, and define the constraints. 

LLMs are refined by human-trained or reinforcement learning, so instead of asking for one suggestion, ask for three or five examples/analogies, and limit them to a paragraph or state a word limit for the answer. 

Large Language Models (LLMs) - a type of generative AI. Popular models include ChatGPT for text output and Midjourney for image output.