8 secrets to great AI prompting

 

1. TELL AI WHO TO BEHAVE LIKE

Often this will just be a description of your role, but it could be another role related to your use case.  Here are a few examples of GPTs that our team has built for internal use at Section. “As a CEO, draft a statement regarding a recent company layoff.” “As a seasoned financial analyst, provide in-depth insights into our quarterly financial data.” “As an HR manager, outline the steps to manage a low-performing employee.”

2. GIVE SPECIFIC REQUIREMENTS ON THE OUTPUT YOU WANT

AI does best with highly specific instructions. Give detailed guidance on what you want your output to contain including tone, format, length, examples, and more. “Create a meeting agenda for our quarterly review. This meeting is 90 minutes long. Include time slots, topics, designated speakers for each topic, and 10 minutes for Q&A at the end.” “Create a list of names for a potential newsletter. Names should be 2-5 words long, focused on marketing strategy, not alliterative, and use a common marketing phrase or pun.”

3. TELL AI WHAT NOT TO DO

Set limitations in terms of both content and logic. This eliminates undesired biases and content, and mitigates the risk of hallucination and misleading information. “Provide recommendations for reducing operational costs, but do not suggest layoffs or reductions in employee benefits.” “I produce baked goods. Write a review about John’s Bagels, but avoid mentioning our own products or making direct comparisons.”

4. ASK AI TO BREAK DOWN THE REASONS BEHIND ITS OUTPUT

Studies show asking to see reasoning leads to better output in general. It also increases AI’s transparency and reliability, because you can see how it got from point A to point B. “Which marketing channels would you recommend to target Gen Alpha? Explain why each channel made your list.” “Give me a ranked list of recommendations for a cloud services vendor. Which factors are you considering in your recommendation?”

5. GIVE AI AN EXAMPLE TO BASE ITS OUTPUT ON

Giving examples allows you to be hyper-specific about your requirements without having to explain them one by one. “Look at this spreadsheet of sales data from the last quarter, then generate a summary similar to this example: ‘In Q3, we saw a significant uptick in sales last August. The data reveals that Product X is our best seller, with a 20% increase in units sold compared to Q2.’” “Create our February monthly marketing email using the attached template. Please use this exact format with the same subheads.”

6. USE DELIMITERS (LIKE HASHTAGS AND SLASHES) OR SECTION TITLES TO DISTINGUISH COMPONENTS OF YOUR PROMPT

AI benefits from lots of context, especially when following an example or template, or analyzing a big block of text. Try breaking up a complicated prompt into sections, demarcated by hashes: ### CONTEXT I am the head of product and want to choose a security compliance vendor. ### OBJECTIVE Examine each of these proposal documents and give each vendor a score from 1 to 10, based on quality, expected yearly costs, etc.

7. TELL AI TO ASK YOU QUESTIONS BEFORE RESPONDING

You want to make sure AI has all the necessary context to answer your initial prompt. Plus, AI is more likely to hallucinate if it doesn’t have enough information. “Before responding, ask me relevant questions to give you additional information you need.”

8. LIST OUT THE STEPS REQUIRED TO COMPLETE THE TASK

Some tasks are complicated, with multiple, dependent steps, and organizing your input helps avoid confusion for the bot. “Compose a project proposal for a new app feature, designed to locate a lost device. Step 1 - Introduce your project idea with a sentence starting, ‘Project idea’   Step 2 - List the objectives of the project with the prefix 'Objectives’   Step 3 - Outline the required resources with the prefix 'Resources Needed’ Step 4 - Estimate the timeline for the project with the prefix 'Estimated Timeline’”