ChatGPT and other generative AI tools definitely have a lot of capability, though getting them to do what you want (the way you want) can be a little challenging if you’re starting from scratch.

Fortunately, there are some patterns and frameworks that have been developed to help people structure their requests (or prompts, in the generative AI vernacular) in a way that gives the AI endpoint (ChatGPT, Bard, Claude, or whoever) the boundaries and guidelines necessary for great output.

Let’s get our assistant assisting!

How do I know how to ask?

There are several ways to explore this question. The best just might be to simply ask a tool like ChatGPT what you need to do to get the best results:

Those are same fantastic tips, indeed. While refinement and conversational styles may help you get where you want eventually, it’s even better if you can be as specific as possible up front.

Through continuous experimentation, several defined prompt syntaxes or rubrics have been developed with simple acronyms that can really help you drive towards the best results.

In this series, we’ll explore some of the popular prompt patterns that I’ve found really useful. Each section will follow a pattern of introducing the prompt syntax or framework, giving an example of a prompt that doesn’t use the prompt format, and then reimagining the prompt with the format.

Off we go!

BAB (Before-After-Bridge)

This simple syntax concept is to use a concise, specific problem statement combined with a goal and instruction. BAB has its roots in traditional copywriting and has worked for decades. The idea is simple–highlight a problem, tease a future where the problem is solved, and insert your product or service as the solution to get the customer there.

BAB can easily be adapted to designing great prompts.

The rubric is divided into three parts:

• Define the BEFORE state or give a problem statement.
• Define the AFTER state or the desired outcome.
• Instruct the assistant to give you a BRIDGE or a path forward.

Without the prompt

For example, let’s say you have a website that isn’t doing well in search rankings. The first interaction is a pretty vanilla prompt and typical of one that someone might enter when trying to figure out how to improve search ranking.

The result is decent and fundamentally sound, also pretty broad and vanilla. Sure, those are all really good ideas, but they’re not really actionable the way they’re written. Of course, not being specific with the exact problem or the exact area led ChatGPT to respond without specificity as well.

With the BAB prompt

Let’s follow the pattern and reframe this interaction with a specific scenario or problem statement, followed by a desired outcome and ending with a specific instruction to provide detailed steps.

Much better! You can obviously further refine this. Some of the suggestions don’t have step-by-step detail, but you could push the model further and ask for expanded detail. You might try prompts like “How do I use Google Analytics to track organic traffic?” or “What are a website’s meta headers? Give an example of an SEO-optimized meta header using my keywords.”

RTF (Role-Task-Format)

In this prompt syntax, you’ll specify the role you want the AI assistant to impersonate, and then provide an instruction to perform and an explanation of how you want it to be formatted or presented.

This syntax has three core parts:

• Assign the assistant a ROLE
• Provide a TASK
• Specify an output FORMAT

Without the prompt

As before, we’ll ask ChatGPT for an answer using just natural language without any framework.

That’s a nice overview, but it lacks the detail necessary to make it actionable.

With the RTF prompt

Now, let’s try again use the RTF prompt format (as this is a GPT-4 prompt with an image, it can’t be shared yet):

A lot more data that you can use–perhaps you could even feed some of those visuals back into ChatGPT (in this case, I used the GPT4 engine, which is also capable of producing images).

Obviously, not perfect (see the underwear with three leg holes), but it’s got some ideas that you can adapt or use as jumping off points for brainstorming.

PAS (Problem-Agitate-Solution)

Like BAB, PAS is a copywriting technique used to present a problem and solution. This time, however, there’s a twist: increasing the pressure. As a prompt framework for generating marketing copy, PAS can be helpful.

• Describe the PROBLEM
• Make the situation worse by AGITATING the consumer
• Position the SOLUTION

Let’s give it a shot!

Without the prompt

In this example, pretend that you’re trying to write copy for a laundry detergent:

It’s kind of boring, just like laundry detergent normally is. It’s not exceedingly compelling.

With the PAS prompt

Now, let’s create a longer prompt that uses the PAS framework:

If you don’t think generated data is creative enough, you can tell the model to modify the temperature.  In terms of ChatGPT, temperature measures creativity, on a scale of .1 (your accountant) to 1.0 (Weird Al on shrooms).

TAG (Task-Action-Goal)

The TAG rubric is another spin on the prompting. Like the BAB prompt, TAG has concepts that translate the thought, “how do we get there from here?”

• Define a job or TASK
• State the ACTION
• Clarify and refine the GOAL

Without the prompt

In this example, we’ll ask for help improving customer service scores.

While these answers are indeed all good things that can be done to improve customer satisfaction scores, there isn’t a lot of detail. Most of these things are pretty obvious.

With the TAG prompt

We’ll rephrase the prompt using TAG, focusing on increasing the Net Promoter Score (NPS).

Once again, using a more specific prompt syntax can help yield more specific results.

Happy prompting!

Series Navigation

More ChatGPT patterns, practices, and prompts >>