It may be hard to believe but there is a near future when your organization will likely be hiring a AI Prompt Architect to build custom applications for your staff and stakeholders. This is something, George (the founder of Whole Whale) has predicted will happen at scale by 2026. Whole Whale has been building AI prompts into custom tools for clients since Q1 of 2022 and we’ve learned a lot about what makes a good vs generic prompt.
Jump to:
GPT and The Grey Jacket Problem
At a Fundraising event in NYC, George was on a panel to talk about AI applications for fundraising and was wearing a grey sports jacket with blue pants. As it happened another person, also named George on the panel was wearing virtually the same outfit.
What’s the connection? Well, both Georges thought their outfits were pretty unique when they bought them and then chose to wear them. When, in fact, it was a fairly common outfit and kind of embarrassing when you end up on a stage in that situation. This is what can happen if your team is using general AI prompts on GPT4, the output will seem unique but it will be an average response that many other groups may get.
When you combine generic prompts with untrained AI’s and then publish it you are walking straight into a Grey Jacket Problem. What’s more, tools like GPTzero.me are getting better at detecting generic AI responses with things like perplexity scores.
Assuming that staff can just wander on to general AI models and create content is like when nonprofits thought it was a good idea to give their social comms to interns a decade ago. Unfortunately, cases like this are already showing up; it was revealed that Peabody EDI Office responded to the MSU shooting with an email written using ChatGPT (February 2023 report by the Vanderbilt Hustler).
The formula for Writing Great GPT Prompts
The best formula for getting unique responses from GPT tools involves a combination of effective prompts, customized settings, adding contextual data, and iterative experimentation. To take a step back, consider how you might give directions to a virtual assistant if you wanted them to write specific copy.
Here is a sample formula to play with for what you’re looking for:
Intent: You are a clever writing assistant that creates emails, articles, and social posts
Context: This AI creates copy like XYZ company that makes paper clips for Millennials.
Brand Personality: XYZ has a fun and helpful tone. XYZ’s founder was originally a comedian.
Data: Here is a list of XYZ’s services and a recent email to customers {data}.
From here you can then make a request for something to be written and then refine it through a narrative or by updating the initial prompt directive.
Here is a breakdown of this process:
- Declare the intent: Ensure that your prompt is clear and specific regarding the task needed. This helps the AI understand your requirements better and provide a relevant response.
- Give Context: Make your style clear by using descriptive language about your audience and flex that thesaurus for tone. For example, you can start your prompt with “Write a heartfelt story for an 8-yr old about…” or “Provide a never-heard-before solution for boomers about…”.
- Add personality: Don’t be afraid to give the AI an attitude or tone. You can even prime it with famous people – write our guide like Abraham Lincoln or Dalai Lama. Experiment with different phrasings, keywords, and other factors to fine-tune your approach.
- Prime with your data: To personalize the AI-generated content, incorporate relevant information from your nonprofit, such as email copy, key metrics, goals, mission and core values. This can also be a good place to add your brand voice or donor personas to describe who you are talking to.
- Fact-check and edit: While AI-generated content can be insightful, it may not always be perfect, especially if recalling facts. These models are presently built on snapshots of internet data and aren’t designed to be research assistants.
- Start from scratch: some interfaces like ChatGPT allow for long threaded chats. These histories can be built on but there is a risk they begin to drift from the original prompt and even forget key elements. This character memory limit continues to grow but there is a limit, and as this Atlantic article points out, GPT-4 Has the Memory of a Goldfish.
When building through the API
For more advanced users, there are ways to better tune models with more variables beyond language. This is where we think the magic really is in developing models that able to reliably creating unique first draft content. Here are couple of our favorite variables:
Experiment with temperature: Adjust the ‘temperature’ setting in the AI model. Higher temperature values (e.g., 0.8 or 1.0) can generate more diverse and creative outputs, while lower values (e.g., 0.2 or 0.3) tend to produce more focused and conservative results. This is available through the API.
Limit response length: Use the ‘max tokens’ setting to limit the length of AI-generated responses. This can help prevent overly verbose outputs and encourage more concise ideas.
Ideas for AI Prompts
Once you understand the formula above you can actually apply it to almost any kind of writing including generating ideas.
Here is a GPT4 response from a tuned model we asked what types of content AI can help with:
- Blog Posts: AI writers can create engaging blog posts that highlight a nonprofit’s mission, success stories, and upcoming events.
- Email Campaigns: AI writers can help nonprofits create effective email campaigns that increase engagement, donations, and volunteer participation.
- Social Media Content: AI writers can craft compelling social media posts, including captions, images, and videos, that promote a nonprofit’s mission and encourage action.
- Grant Proposals: AI writers can help nonprofits create persuasive grant proposals that secure funding and support for their initiatives.
- Newsletters: AI writers can help nonprofits design and write newsletters that keep donors and supporters up-to-date on their latest activities and achievements.
- Press Releases: AI writers can help nonprofits create attention-grabbing press releases that generate media coverage and raise awareness of their cause.
- Video Scripts: AI writers can help nonprofits create compelling video scripts that showcase their work, impact, and vision.
- Annual Reports: AI writers can help nonprofits create engaging annual reports that highlight their achievements, financials, and future plans.
Just because we don’t want to completely be replaced by a large language model, here are some more clever uses we have played with:
- Brainstorm a topic, ad copy, or actions like how volunteers can help X. Give the context, audience, and guardrails described in the above formula and ask for 10+ examples to increase the variety. Pro-tip here is to increase temperature setting to increase randomness and decrease the presence penalty.
- Create a donor persona chatbot and then have a conversation with it and test what it thinks about marketing messages in a role-playing game.
- Build a model trained on your founder and then have it summarize articles with opinions. We did this with George, and then he interviewed the AI version of himself, we think it scared him.
- Build large training guides on a given topic. When doing long-form content, begin by co-creating the outline of the topic and then work to fill it in by resubmitting the outline interactively.
- Build a tool for your audience that helps with a task. Here is an example of how Nonprofit.ist, a trusted nonprofit expert network used it to build an AI RFP generator.
- Build out code (see the bottom).
Conceptual Prompt Narratives
To avoid generic or unstructured prompt responses try using the Problem, Agitation, Solution (PAS) framework to build an AI copywriting tool. It is also an effective way to structure outputs and can be used for various types of content, including sales pages, emails, and blog posts.
The PAS framework consists of three parts:
- Problem: The first part of the framework is to identify the problem or pain point that your target audience is experiencing.
- Agitation: The second part of the framework is to agitate the problem by highlighting the consequences of not solving it. This involves creating a sense of urgency and making the problem feel more significant, pressing, and frustrating.
- Solution: The final part of the framework is to present the solution. This should be a concise explanation of how your product or service can solve the problem and help your audience.
As an example we used the PAS framework and asked GPT4 to apply it to PowerPoetry.org:
Problem: Many young people struggle to find a safe and creative outlet to express themselves through poetry, especially those from underserved communities.
Agitation: Without access to a creative outlet, young people may feel unheard, invisible, and may suffer from anxiety, depression, or other mental health issues. They may also lack the support and resources they need to grow their skills and share their work with others.
Solution: PowerPoetry.org is a nonprofit organization that provides a safe and supportive community where young people can share their poetry, connect with other poets, and grow their skills through free online resources and events. By supporting PowerPoetry.org, you can help provide a platform for young poets to express themselves, connect with other artists, and gain confidence in their abilities. Your support can also help ensure that underserved communities have access to resources and support that can help them thrive.
This example shows how the PAS framework can be used to create a persuasive appeal for PowerPoetry.org. By identifying a problem that resonates with the audience, highlighting the consequences of not solving it, and presenting a clear solution, PowerPoetry.org can create compelling content that motivates supporters to take action. Other examples of the PAS framework in the context of PowerPoetry.org could include addressing issues such as censorship, lack of artistic representation, and low literacy rates, and presenting solutions such as promoting free speech, supporting diverse voices, and providing literacy resources.
Current AI Models
There will be a proliferation of LLMs – large language models trained on neural networks that have been fed billions and trillions of text data across the web. They may become more specialized and more built into native tools like word processors, websites, and email interfaces. What will still remain is the skill needed to get unique, personalized output so you don’t end up with the Grey Jacket Problem.
- GPT4 and ChatGPT – https://OpenAI.com
- Google Bard and LaMDA – Google Bard
- Claude – Anthropic | Introducing Claude
Beginner Code Prompts
Up to this point, we have only focused on text copy-based prompts, but the power of this tool goes way beyond text completion. Your developers should also be playing with code completion and tools like copilot from Github. Note that all code should be reviewed by someone who understands the language before blindly being tossed on a site.
For code here is a rough prompt approach:
- Define the user experience desired
- Explain the code bases you are using or need it to be built it
- Refer to the library or copy relevant API documentation
- Note that you want it also explained in pseudo-code
Here is an example of a donation thermometer prompt we made for fun with GPT4:
It is also possible to have it build things like SQL queries:
If you’re curious about what an AI prompt training or a purpose-built series of AI tools could do for your organization, you know where to find us!