Overview: Interview Prep Bot
This bot was created for Voiceflow’s PromptHacks hackathon on Apr. 28-30, 2023. The challenge was to build a creative and practical project using GPT-4 and other LLM models over the course of a weekend and submit the project and a demo video for judging.
Role: Conversation designer
Collaborated with: 2 other conversation designers.
Tools: Figma, Voiceflow, ChatGPT, GPT-4
Time: Apr 28-30, 2023
Problem
5.8 million people were unemployed as of March 2023, or 3.5%, according to the U.S. Bureau of Labor and Statistics.
Solution
An interview prep bot that helps people prepare for interviews and improve their answers to interview questions.
(The bot is in the style of Snoop Dogg for the demo, but you can switch out the persona to anyone you want!)
Demo Video
Process
Before we started building the bot, we wanted to scope out potential use cases to make sure we were choosing the best one for the challenge.
We were very ambitious and started out with broad use cases that overwhelmed us. We had a hard time narrowing down our ideas, and figuring out how prompt-chaining would solve that issue.
If you want to take a closer look at our strategy process, here is the link to our FigJam strategy board.
We went back and forth between 3 ideas, but eventually we honed in on one idea that we felt good about: a career coach bot that could help people create and improve their resumes and cover letters, potentially generate an AI headshot, and help people prepare for interviews.
We quickly realized that we were way too ambitious in the amount of time that we had. Without a developer and more time, we weren’t going to be able to have the right integrations ready in time.
After going back and forth for way too much time, we started building out the interview prep portion of the bot. We felt it was the most unique feature, it provided the conversational aspect that is missing from interacting directly with ChatGPT, and it would actually be useful for job seekers.
Here’s the flow we ended up with:
It may look like a simple flow, but we had our fair share of issues trying to work out all of the problems we ran into.
Getting the prompts correct so we got the output we wanted was our biggest challenge. We spent hours trying to write, test, and rewrite each prompt, and explore which models worked best with each prompt (we had the options of GPT-3, ChatGPT, and GPT-4).
We also played around with the temperature and tokens (the temperature affects the way the bot responds and the token affect the length of the response) and did a lot of testing to see how that affected each response.
We also found that the responses were long-winded and tended to ignore conversation design best practices such as asking multiple questions at a time and asking questions in the middle of long paragraphs. We realized that with a persona, it was much easier to digest those long-winded answers and immediately became a lot more fun for us to design. Here’s an example of the difference:
We also had persona issues. We tried writing out a few different personas at first, but we ended up using Snoop Dogg for the demo. However, because we had experimented, it seemed like some of the personas we had created before were still carrying through the middle of the experience, so the bot personality kept changing throughout the conversation.
Our last major challenge was making sure our conversation loop was working correctly. We spent a few hours testing and revising, until we were satisfied with the input and the output of the bot. We found that any tweak we made, no matter how small, affected the entire experience, which is why we had to spend so much time testing and making sure any change we made didn’t affect anything else in the bot.
While we had our ups and downs throughout this hackathon, we all learned so much and had a lot of fun doing it!
If you want to play around with the prototype, here's the link:
Here are a few quotes we received from people that tried out our bot:
“As a former career coach, I can think of so many people that would use this!”
“What a terrific idea. And fun! Your bot works like a champ! Well done!”
“I’ve coached my fair share of candidates preparing for interviews, but I promise you, I was never as cool as this bot with a Snoop Dogg persona. This was too much fun! I can think of sooooo many people who would want to use this bot.”
Next Steps
Our initial vision for this bot was a full-fledged job search experience that would handle generating a new resume and cover letter, AI headshot, career advice, and helping a user upskill. In the short term, we’d like to add an ASR component so that the user can speak their response and interact with voice. This would also allow sentiment analysis and the ability to provide more meaningful feedback.
Eventually, we would love to build this out further and do some user research to see if this is something people would actually be interested in. We would also love to play around more with the integration of ChatGPT, GPT-4, keeping the persona consistent throughout the experience, and experimenting with when to use memory vs. prompting or both.
