GPT-3 Prompt Chains are Transforming the Way We Work
I discovered GPT-3 exactly one year ago. My original post on the Open AI forum today is dated Feb 12th 2022.
I remember telling people it reminded me of what Facebook felt like when I first discovered it. Back when it was just students who had university emails. That feeling of technology that was going to change the world, it was exciting and brimming with possibility.
I couldn’t understand why no one was talking about it. I didn’t understand why every student I know wasn’t using this. How had it not already transformed the business landscape? I went to the Open AI forum to find examples of people who had transformed their lives with this tech. I was disappointed. I thought I was late to the party but it turns out I was way too early.
A year later and ChatGPT has exploded on the scene. My Twitter feed has been a non-stop parade of interesting prompt screenshots and entertaining one-off use cases. The use case examples that come with GPT-3 and ChatGPT are fantastic.
But very few stories about people who have successfully augmented their work or lives with this technology. So I put together the best stories I could find for you. This post is what I wish I found on that forum a year ago.
Product designer turns customer interview insights into jobs to be done.
Shavin Peiries is a product designer who used to run a design agency called “Very Bad Wizards”. They focused on helping corporate businesses apply design thinking to projects.
Analyzing customer interviews is a slow, manual process
A large part of Shavin’s work involves interviewing customers to understand their problems. This involves speaking to people, transcribing the conversations, highlighting insights, organizing them into themes, and then synthesizing the findings. It can take days to analyze a batch of interviews because the work can be so subjective.
Synthesizing customer insights in half the time
Shavin figured out a simple chatGPT workflow that organized all the highlights from the interviews into themes.
INPUT PROMPT 1: Group these customer quotes into crisp, actionable themes. Include the number of times a quote appears for each theme, who said the quote, and a snippet of the quote relating to the theme. Do it in the table format.
Then Shavin had the brilliant idea to get ChatGPT to rewrite the themes in the job story format so that they were easier for everyone to understand.
INPUT PROMPT 2: Now could you write them in the job stories format starting with When I...?
The final step involved identifying the most important customer challenges his clients would need to address. Shavin asked ChatGPT to take the customer quotes and create problem statements in the “how might we” format to reframe problems in the previous step as opportunities.
INPUT PROMPT 3: Generate 5 problem statements in the HMW format so that we can be discovered by these types of customers
ChatGPT successfully turned a collection of quotes into a set of key customer challenges that Shavin’s client could then brainstorm solutions for. All in half the time it usually takes. Plus, the opportunities it came up with were at the perfect level of abstraction to get ideas flowing, not too narrow nor too broad.
Stuff to watch out for
- Regenerating the same prompt multiple times leads to a different output. As Shavin puts it, “For example, a quote filed under “personal satisfaction” might be grouped under “desire for ownership” on a different, subsequent run. This lack of consistency can make it difficult to use effectively in a team setting.” If you are looking to recreate this use case I’d suggest using the Open AI sandbox rather than the ChatGPT so that you can turn the temperature setting down low to avoid this problem. Temperature corresponds to creativity, and the lower it is the more predictable the output becomes.
- The other issue was with the number of quotes you could feed ChatGPT in one go. Rather than giving it ten thousand quotes to work with, Shavin had to feed the quotes in 40-50 at a time. To overcome this limitation, keep reading because the third use case in the post covers a brilliant workaround for working with lots of text.
- You also have to watch out for when GPT-3 hallucinates information. For example, in the second step, ChatGPT wrote a job story about “not being constrained by the landlord” but there was no mention of a landlord in the quotes or interviews.
- For a full write-up of this case study, along with all the output for each of these prompts, check out Shavin’s guest post on the Buildspace blog.
- If you’d like to know more about Shavin Peiries, you can read about his work on his blog or you can get in touch with him on Twitter.
Ahrefs’ VP of Marketing uses AI to Craft Compelling SEO Content
Sam Oh is the Vice President of Marketing at Ahrefs. He recently released a YouTube video showcasing the best ways to utilize ChatGPT in search engine optimization (SEO).
PROMPT INPUT 1: Write 10 click-worthy titles for my blog post on ChatGPT for [your topic]. My working title is [Your current title]. Front-load the keywords "[keyphase you are targetting]"
You can then adjust the results by requesting that it reword the titles in a specific style. For example, in Sam’s case…
PROMPT INPUT 2: Make these titles sound like a Mr. Beast video
Once you’ve selected your preferred title, ask for a brief outline of the post.
PROMPT INPUT 3: Write an outline for a blog post about [insert title]
When it comes to the topic, ask ChatGPT to address frequently asked questions or the most widespread misconceptions. Having an outline to revise and incorporate your expertise into is an excellent method to overcome writer’s block.
Contrary to popular belief, ChatGPT is not ideal for writing long-form articles. Although it can write complete articles and even imitate a specific author’s style, there is no advantage to writing generic content for SEO purposes. Everyone now has access to ChatGPT. If people click on a search result looking for information, it is critical that your article provides the expertise and detail they need. If you use your expertise to write the article and treat ChatGPT as an assistant, you won’t have to worry about search engines penalizing AI-generated content in the future.
Note that ChatGPT can still be a valuable tool for proofreading once you have finished writing.
PROMPT INPUT 4: Proofread this: [insert paragraph]
Once you’re ready to publish, ChatGPT can help you create your meta descriptions.
PROMPT INPUT 5: Write a meta description thats no more than 156 characters for a page titled: [insert title]
If you found these prompts useful, be sure to watch the whole video as it covers a range of other use cases, including generating feature snippets, writing Google AppScripts to find email addresses, creating schema markup, and performing regex searches in the Google Console. Most importantly, the video explains why using ChatGPT for keyword research is not recommended.
Journaling with GPT-3
This use case has to be my favorite so far.
Dan Shipper, from the beautiful publication “Every”, started using GPT-3 to journal as a personal development practice. Rather than staring at a blank page, he wanted journaling to feel more like a conversation. He would ask GPT-3 to pretend to be someone and then ask him questions to help unpack things in his life. More like a mashup of journaling and talking to a friend.
He started experimenting with prompts like this:
You are Socrates, please help me with an issue in my life. Please ask me questions to try to understand what my issue is and help me unpack it. You can start the conversation however you feel is best. Please end your responses with /e
Michelle Huang’s Twitter thread went viral when she pushed the idea further by feeding GPT-3 her journal entries so that the person she was talking to was a younger version of herself.
Michelle asked her younger self what she thought about the world, what her values were, and what her perspective on things was, stuff that present Michelle was grappling with. Michelle even asked her younger self if she had questions. They talked about whether she had followed her dreams or not and if she was happy with her life at the moment. She was reminded of the parts of herself that stayed constant through the years, but also of the parts she forgot or buried as life went on.
The prompt she used to do this was:
The following is a conversation with Present Michelle (age [redacted]) and Young Michelle (age 14). Young Michelle has written the following journal entries: [diary entries here] Present Michelle: [type your question here]
One of the problems with this approach is the technical limitation on the amount of text GPT-3 can compute in one go. The latest models have a maximum limit of 4097 tokens, which is about 3000 words. If you include 2000 words of journal entries in your prompt, you can only have 1000 words of conversation before it stops working. Not ideal if you’ve been journaling for over 10 years.
Dan found a way around this memory problem by using a tool called GPTIndex.
GPTIndex can store 10 years of journal entries by breaking them down into smaller chunks, which are then indexed so they are easy to find and summarize. When you ask a question, the tool finds all relevant chunks and uses them as context for the conversation, effectively bypassing the 3000-word limit so you can work with as many journal entries as you want. You can read the full article, “Can GPT-3 Explain My Past and Tell My Future?” for more specifics on how he pulled this off.
And it gets stranger.
With the ability to feed GPT-3 years of journal entries, we effectively creating a “second brain” that we can talk to. But there’s nothing stopping us from creating a similar experience with someone else’s work. For example, Dan loves listening to the neuroscience podcast “The Huberman Lab”. Each episode is several hours long and it’s a hassle to scrub through hours of audio when he has a specific question in mind. So, he transcribed all the episodes. When he has a question, the bot finds the most relevant sections of the transcripts and sends them to GPT-3 with the following prompt.
Answer the question as truthfully as possible using the provided context, and if the answer is not contained within the text below, say "I don't know." [ relevant sections of Huberman Lab transcripts ] Q: What is task bracketing? A:
You can see the kind of pinpoint accuracy it comes back with in this demo on Twitter, or you can read about the who experience here, “I Built an AI Chatbot Base On My Favorite Podcast“
Dan repeated the process with Lenny Rachitsky’s entire archive of content. Lenny has been writing and talking about building software products and growth for years. Indexing his expertise and using it as context for questions allows GPT-3 to produce answers with specialized knowledge and nuance in a way that vanilla GPT-3 can’t match.
As Dan points out, this could give people the ability to monetize the content they’ve already created in new ways: “A new class of content creators will learn to create compelling chatbot experiences that combine their personality and worldview for their niche audience in the same way that some creators learned to create compelling YouTube videos, newsletter articles, or TikTok clips.”
We went from journaling to therapy, to building your own second brain, to creating what is probably going to become an entirely new form of media. And we’re still just talking about ingesting one person’s body of work here. Copyright complexities aside, what happens if I want to feed GPT-3 multiple bodies of work from people I admire and respect?
These are strange and exciting times.
I will continue to delve into this rabbit hole and share what I find.
If you want to create your own little chabot you will need someone with programming expertise to set it up for you. But you can use Michelle’s or Dan’s initial prompts to start journaling with ChatGPT.
If you have an AI use case or find an interesting story online, please send it to me on Twitter @joshpitzalis.