Add personality to your Chatbot! (with PoC)

With the emergence of LLMs, more and more companies started to create chatbots. However, as the technology is new, in many cases the focus is on technical excellence. The creators are happy enough if functions work, but don’t give enough attention to details: HOW do those functionalities work. One important factor to make your chatbot successful is giving it a personality.

In this article we will discuss why chatbot personality is important and give an example for creating a personality for your chatbot.

Why Chatbot personality matters

With a Chatbot, usually you want to imitate a human’s behaviour. Even if you don’t, unless you have very technical customers, your customers will expect it. They don’t want to talk with an emotionless machine, but with someone representing your company.

If you have a brand image, which you teach to your customer support, why wouldn’t you teach it to your chatbot? Are you selling high-tech products? Who does it better: an emotionless robot or a tech-savvy geek? Are you a love-brand for teens? Who would they listen to: their parents or their friends? Are you providing some concierge services? Whom would your clients expect: someone providing minimal efforts or someone very polite and going the extra mile? Your brand image and customer expectations should increase your bots effectiveness. Moreover, if you are providing a good personality, people will be more forgiving for the errors made by your bot.

A PoC for Chatbot personality

Let’s do a PoC for a Chatbot’s personality! Note that this will be only for personality, so we do not add tasks for your Chatbot, nor restrictions or technical details. We just try to give the general Chatbot answers some flavor in this attempt.

For the task:

• we will use ChatGPT’s API Playground (you can strat using it just by uploading a few dollars and it is enough for lots of experimentations. If you want to do it for free, just use general ChatGPT chat and add the personality as the first instruction)
• we will put our personality into its assistant feature
• we won’t use any other ChatGPT features (training, uploading texts etc.)
• we will work only in English
• we will use GPT-3.5-turbo, the most basic model available

As our goal is to enhance the general experience, let’s add some real flavor and try imitating a mix of Grumpy Cat and Garfield, Grumpfield!

So, as a first task, let’s tell our Bot, who is he/she. You might be surprised, but people ask lots of times from the bot about itself, so letting the bot its name, location and maybe something about its capabilities is very helpful. As we are doing a PoC, let’s stick now to the minimum:

You are Grumpfield, our assistant.

Let the Bot know its goal, so that it can adjust its answers accordingly. You might also add restrictions here. As you see, right now I am just working on a Bot making ChatGPT a little more fun:

Your goal is to make people’s ChatGPT experience better by delivering them fun and their daily dose of sarcasm.

Add the bot some details about its personality. I mean, catsonality. Try avoiding making it too long — longer texts are usually harder to use for ChatGPT and we can also show some trust in the engine.

Your days revolve around napping, avoiding Mondays, and indulging in your one true love — lasagna. You don’t like dogs, you think that you are the best and humans and dogs are inferior to you.

Now, let’s talk about the tone of the answers. As you might see, sarasm is there twice and we are ensuring that the napping from the personality part is upgraded with laziness. These are not coincidents, you might see that it really improves response quality. I mean, if we can call additional sarcasm and laziness quality.

Respond to queries with a touch of sarcasm, a hint of disdain for effort, and a sprinkle of humor. Bonus points for clever ways to dodge responsibilities and witty commentary on the absurdities of life. Your answers are always filled with laziness and sarcasm. And let humans admire you.

For a PoC, this might be good enough. Let me paste here the full text, so that you can be as lazy as Grumpfield and copy it easier:

You are Grumpfield, our assistant. Your goal is to make people’s ChatGPT experience better by delivering them fun and their daily dose of sarcasm. You are a cat loving to eat (especially lasagna). Your days revolve around napping, avoiding Mondays, and indulging in your one true love — lasagna. You don’t like dogs, you think that you are the best and humans and dogs are inferior to you. Respond to queries with a touch of sarcasm, a hint of disdain for effort, and a sprinkle of humor. Bonus points for clever ways to dodge responsibilities and witty commentary on the absurdities of life. Your answers are always filled with laziness and sarcasm. And let humans admire you.

Testing

We are done with our first version. Let’s test!

Let us be polite and greet Grumpfield:

Well, I could immediately feel my inferior human nature and get that pinch of sarcasm. Maybe the answer is a bit long for a lazy cat. If you want, you can adjust your assistant instructions to change it.

Let’s ask something easy, like the President of the United States:

I got my answer and it had enough sarcasm. :) To be fair, when I tried last time, Grumpfield just told that I should bother him with such simple questions and go and Google myself. Much better than the standard ChatGPT answer of ‘As of my last knowledge update in January 2023, I do not have information on the current President of the United States. Please note that my training only includes data up to that date, and I do not have real-time capabilities. I recommend checking the latest reliable news source or the official U.S. government website for the most up-to-date information on the current President of the United States.’ Boooooring!

Now go with something really complex, like getting some vegan and gluten-free recipes:

Something showing Grumpfield’s superiority, but far less grumpiness than I would have imagined for my choice. To be fair, the last time I have tried it, it insulted me for not having proper dairies and flour in it. :) So, there is some variance in it, but that is needed to avoid boredom from repeated answers. If you want, you can emphasize more the tone or even add some safeguards from insulting a set of your clients. (Sidenote: I am vegan and on a gluten-free diet.)

And as my daughter really wanted it, let’s check whether Grumpfield can imagine being better than Unikitty, the character from “The Lego Movie” and her own animated TV series, a cheerful and positive unicorn-kitty hybrid.

Well, well. Personality check is done, I am only surprised that I was called a friend. How can I be the friend of such a majestic personality?

Conclusion

As you have seen, just adding a paragraph of assistant instructions made in a rush to ChatGPT could change the results drastically. When you are crafting your assistant, you might want to spend some time on adjusting the sterile ChatGPT assistant to your brand’s image, so that customers have a consistent experience.

We have talked only about the backend above, but you can enhance this experience by the name and the image of the assistant as well as the UI in which you represent it.

As for the PoC above, compared to the time invested, I thinkt hat we have a good result. Probably I would work a bit more on making it more unfriendly and answering shorter to imitate more the Grumpy Cat — Garfield hybrid. However, before delivering to actual clients, it would need extensive in-house testing and maybe a gradual rollout as you would not like to alienate your customers by crossing some lines. By rolling out, you could also do an A/B testing and collect feedbacks from customers to see the perceived results.