Sharpen your ChatGPT / AI Market Research and Insights Prompts with the FEAT Framework
A simple framework for extracting maximum value from your prompts.
Truly investing in your ChatGPT / AI prompts is the best way to achieve consistently exceptional outputs.
As the old saying about Generative AI Prompting goes: "Garbage in, surprisingly decent quality out." That's fine if you're seeking a recipe for tonight's dinner or travel itinerary inspiration. But when it comes to applications in the professional services arena, where the stakes are very high, getting extraordinary results requires extraordinary prompts.
I created a simple framework to help you get the most out of your prompts: FEAT Prompting. This framework is the result of a year of prompting experience, and hopefully it helps you accelerate your AI adoption curve. Below is an overview of the approach, with a few brief examples to help you as you refine your prompting craft.
You've seen the headlines where Prompt Engineers, a role almost completely unheard of this time last year, are already commanding $300k+ salaries, and there's a good reason for that: prompts need to be delicately and intentionally constructed to ensure the output is maximally relevant, and maximally useful.
A Shift in Workflows
One foundational shift in how tools like ChatGPT are used versus Google/Bing/search engines is how much time is spent on the input/prompts.
Previously, the trend with search was towards shorter, punchier search terms in order to most effectively (and speedily) get the desired results. Your search terms are entered in a couple seconds, and you're on your way to click through the results.
However, with Generative AI, taking time to think deeply about designing your prompt and building it out in a structured way will yield much more effective results.
Instead of a quick one-line prompt, I will spend 15-30 minutes building and refining an important prompt. Not only does that provide exponentially better results, but it is also an investment in future prompts, as I make sure to save the work in my prompt library (more on that another time…).
Starting at the End
Before putting pen to paper (or, I suppose, fingers to keyboard…), think critically about what output you need, how it will be used, and what format will be most useful. The more you can put your mind in that future state, the better you will be able to build out the prompt.
Once you've paused to really think that end goal through, only then is it time to build out your prompt.
Introducing the FEAT Framework
The framework functions a bit like a funnel, starting with guiding the model as to which 'corner of its corpus of information' to focus on. From there, each step refines the goals, objectives, and instructions. The four key components are:
1. Framing
2. Expectation
3. Assisting
4. Task
When building out the prompt it's important to remember that, in addition to guiding the effectiveness and quality of the output, you are also helping craft the UX/UI for your interaction or the interaction or GPT you are building.
Along those lines, it's also important to know that this FEAT framework can be used for direct prompting (e.g., using ChatGPT) or for building out your own GPTs that you or others may use.
1. Framing
These models are trained on virtually every type of data: creative fiction, social communications, academic journals, artistic creations, business theory, etc. In order to get the model to "focus" on the areas most relevant to your goals, the prompt can help maximize its ability to add value while being especially relevant to what you are looking to accomplish.
The easiest way to start to frame your prompt is by simply telling the model "who to be".
For instance:
"You are an expert qualitative moderator in the market research and insights industry, skilled at enabling creative and unique conversations with participants that are engaging while efficiently addressing the business objectives of an engagement. You have deep experience with consumer packaged goods, retail and ecommerce channels, and product innovation and optimization."
2. Expectation
Next, you need to articulate the purpose and goals of the interaction. (Note: I do recommend thinking of it as an interaction and not just a one-off prompt, as I've gotten the best results by continuing to engage in "dialogue" with the model) This both informs the overall output, as well as how the model will interpret the remainder of your prompt.
The nature of the ask:
Are you looking for a clearly defined/formatted output? Are you looking for help with creative brainstorming? Do you need the model to role-play with you to help practice an important meeting you have coming up? There may be more than one desired output, but setting the stage upfront will help you arrive there more efficiently.
The nature of the interaction:
Do you want the model (and output) to be funny and conversational? Or professional and precise? You should proactively set the tone from the start.
The nature of the flow:
Do you want the model to automatically give you its best attempted output? Or do you want it to ask clarifying questions as it goes? Is this a singular output or are there multiple stages to the interaction?
3. Assisting
Now it's time to assist the model by providing any important pieces of context. (Note: This is all with the obvious caveat that you shouldn't provide any IP/confidential information)
I like to use a funnel approach here as well, going from macro to micro:
What is the specific industry or market? What information about the current state or anticipated future state of the market will help?
What key information about the brand or product would be helpful?
What is the specific objective of the task-at-hand? Who will be using the output? How?
What hypotheses or theories should the model take into account?
What other details are important to factor in?
Put in as much detail as you feel is necessary to guide the results, but also think holistically about what is "necessary" to guide to make sure you don't inadvertently bias the output in an unwanted way.
4. Task
Only after the above context has been provided is it time to provide the specific task for the model. This component could have a whole book written about it given the infinite possibilities, but I'll distill the key components.
Define the objective:
What does the end output entail? Is it an outline? Is it imagery? Is it a block of text?
Is it finite (a singular goal) or ongoing (continuous open dialogue over time)?
Are there different steps or stages to the approach? What are they? How do they relate to each other?
Define the output:
What is the format of the output? Is it in paragraph form? Bullet or numbered list?
How should it be organized? Are there headings? How should those be created?
How long should the output be? A list of 20? A 500 word blog post?
Should references and links be included?
Do, Learn, and Repeat
The above examples are broad, but will hopefully help you with framing your prompts to yield maximum quality outputs and impact.
The level of detail you include will vary proportionately with how critical the output is. For 'mission critical' output, my prompts can get extremely detailed. But I've found the time invested in thinking it through yields a multiple on time saved (and the impact) in the end.
You'll get a feel for how the models react to different components, so take the time to experiment as well and take note of what works best for you (saving that prompt text for future use).
--
Go forth and try applying this framework to bolster your prompts. And please let me know if you have any questions or other thoughts!
Onwards,
Matt Walker