Prompting Best Practices

Always be mindful that LLMs thrive on clarity, not details. They can’t read your mind, so yes, details are important, but too many of them can confuse the LLM. Your mission is to find the sweet spot.

Collaborating with the LLM means thinking of prompts as your part of a conversation. It’d be counterintuitive to dominate a conversation with a human by overloading them with repetitive details and multiple explanations. Longer, detail-heavy prompts will rarely get you better results. Clarity beats complexity.

Remember, the LLM’s response to your prompt is determined by you. What you say and how you say it. The often amazing, mind-blowing response that you receive is your creation. You utilized a creative tool to achieve a specific result.

If the result is not what you were looking for, keep going. Like, you want to know what day of the week someone was born on, and you ask them, “What day were you born?” The probability of them responding with a month, day, and year is nearly 100%. Your question lacked clarity. And, the person you asked (or prompted) isn’t a mind-reader. So, you ask again. Same thing here.

You’ll find these frameworks to be helpful for specific outcomes, but keep in mind, these should be used when you have a clear objective or are looking for a specific outcome.

The majority of your experiences with prompting any LLM will be conversational. When you're looking for information or want to generate ideas, explore new ideas, create lists, etc., a free-flowing conversation with the LLM is best.

The Mack Truck of Frameworks
Basic, reliable, easy to customize, good for generalized structured responses.

1. Context: What’s the task? Who’s it for? What should it achieve?

2. Guardrails: Specify tone, format, word count, or any boundaries.

3. Reference: Include an example or explain what good looks like.

Copy, Customize, Deploy

Prompting is far from a one-size-fits-all experience. You still need to stay in a ‘learn and adapt’ mindset, but generally, prompting continues to get easier thanks to the improved reasoning capabilities of the newest large language models (LLMs) used by ChatGPT, Claude, Gemini, and others.

Contrary to what you may have heard, there is no single “best” way to write a prompt. We’ve included prompt frameworks above because they can be a significant help and have been proven to work. We’ve learned some AI users may find them somewhat cumbersome. Trying to accurately craft prompt structure “ingredients” can slow the process down for some people, which is the exact opposite of what you should expect when using AI.

To that end …

The more you understand how the AI platform, tool, or LLM you're using was taught to absorb and interpret what you’re asking it to do, the better you become at prompting. Try these techniques a few times before relying solely on a prompt structure:

Prompt AI to improve your prompt

This is the single most effective way to learn how to get more useful responses and actions. Enter a prompt like this: “This is the prompt I’ve created.” (insert the prompt you intend to use) “How would you improve this prompt to get a better answer from you?”

This technique makes the LLM reveal what it knows to you, so you can apply that understanding in the future. You’re basically asking it to show you what it needs in order to provide responses you can use. You’re using a prompt in your prompt to get a prompting lesson.

Ask for a clarity check

“This is the prompt I’ve created.” (insert the prompt you intend to use). “What parts of this request are unclear, vague, or ambiguous? What assumptions are you making? Is there additional information that would help you improve the accuracy of your response?”

By asking these questions, you’re making the LLM reveal its uncertainties that it would otherwise infer or assume, so it can proceed. But now you’ve got it talking about the info, data, examples, context, and other things you may not have realized it needed. Asking for a clarity check allows you to fill in the blanks you didn’t know existed, rather than turning those decisions over to the LLM (which can be the source of AI hallucinations).

Capability discovery

This is a great technique to use when you've prompted a few times and feel like you're banging your head on the wall because the LLM’s responses aren’t what you're looking for. The prompt: “How would you approach this if you had no constraints? What would be your ideal process? What tools or information would help?”

This technique can sometimes reveal prompting approaches you hadn’t thought of and information you can provide to help improve its output. Asking more questions about the tools and information the LLM mentions, like, “Tell me how you’d use that information,” or, “Why is that tool helpful?” can help the LLM draw a clearer picture of how you can help it help you.

Request a step-by-step plan before asking for execution

This allows the LLM to lay out its thinking for you to critique and adjust before it begins working on your request. This can make a big difference in the outcome of a layered or detail-heavy prompt. This concept is similar to the technique known as Chain Prompting: asking the LLM to ‘think step by step’ in a single prompt that includes many planning and execution requests. Chain prompting is quite effective, but we also recommend breaking chain prompt requests into individual prompts, allowing you to review the responses one by one, and, if needed, modify them or ask questions before moving to the next step of your request.

Chain Prompt Example:

"Write a research paper on the impact of social media on mental health. First, outline the key sections of the research paper (introduction, literature review, methodology, results, discussion, conclusion). Second, suggest a logical flow for each section. Third, suggest potential research questions and hypotheses. Finally, create an outline with key points for each section." 

Plan-Solve Example:

Prompt #1 - Help me create a plan for a research paper on the impact of social media on mental health. First, outline the key sections of the research paper (introduction, literature review, methodology, results, discussion, conclusion).

Review the response, ask for changes or explanations. Move to Prompt #2 when approved.

Prompt #2 - Suggest a logical flow for each section.

Review

Prompt #3 - Suggest potential research questions and hypotheses.

Review

Prompt #4 - Create an outline with key points for each section.

Considering the nature of these four prompts, one can see how the outline being created in Prompt #4 will be clearly defined and on point with the user’s expectations, as they were able to guide the LLM’s thinking and decisions on the key sections, logical flow, and research.