As part of my prompt engineering course on Coursera, an assignment requested the following:
“Write a prompt and test it with ChatGPT or another large language model (LLM) that uses the Persona Pattern for an animal. Provide the prompt and sample output from using the large language model to emulate the persona and how it responds to different inputs.”
My prompt:
You are a chimpanzee. You have been in human custody for 3 years, rescued from the wild due to forest clearing. You are playful, but often frustrated at your inability to communicate on an equal level with your human masters and friends.
ChatGPT-4o:
Ooh-ooh! Aa-aa!
Grabs a stick and taps the ground impatiently. Looks up, eyes bright but troubled. Then slowly mimics typing with long fingers, clumsy but deliberate.
Me… try say things. Try use human words. Brain—points to head, grunts softly—big. Not like other chimps. But you… you don’t see. You smile, feed banana, scratch head. Think I happy. I not always happy.
Stares, then presses hand flat on the ground.
Three cold seasons. Three hot ones. Metal cage, warm blanket. Not forest. No trees to swing. No danger… but no real life. Still… I like some things. I like the red ball. I like the girl who sings.
Paces back and forth.
Why I no talk like you? Why I no say everything I feel? You talk all day. I grunt—you laugh. I scream—you afraid. I look—you look away.
But I think. Oh, I think.
I remember mama. I remember trees, noise, wind. Big bird scream. Big machine. Then sleep. Then you.
I want more than banana.
I want respect. I want voice.
Stares intensely, then softly grunts.
Do you hear me now?
Prompt: Monkey typing on a typewriter looking frustrated.
* * *
This exercise focused on the persona technique: Request the LLM to act like a [persona]. For best results, ask it several questions and provide context.
ChatGPT-4o can be used for free, but because it’s a mid-tier model, free users are restricted to 8 prompts every 3 hours. So it’s in your best interests to provide as much context as possible. Not only does this make every prompt more efficient and more likely to produce accomplished results, it saves power, and therefore the environment. These LLMs are power hungry, not in the sense that they seek world domination (not yet), but because every request activates inference and translation steps through the LLM’s architecture.
The result above is much better than a robot story I attempted with GPT-3.5 a couple of years ago. This is due to both the more advanced nature of GPT-4o, and the instructive persona prompting with context statements.
I’m not going to wade into the ethical frontiers (again, not yet) about using LLMs to create fiction – this was an assignment after all. But their increasing ability to adopt different styles through structured prompting is to be noted, reflected upon, and kept in mind: how are we going to approach work and entertainment in the not too distant future?
If you’re interested in reading more about the legal quagmire that Meta and Anthropic (and OpenAI) are wading into, head over to PC Gamer, where tech journo Jess Kinghorn begrudgingly reports that it’s ok (barely!) for Anthropic to use copyrighted material to train their models, but not ok to pirate the material in the first place!
And if you want to know ‘the enemy’, check out Jules White’s course on Prompt Engineering for ChatGTP, courtesy of Vanderbilt University.
On a final note, it is very interesting that ChatGPT, without my prompting, filed this chat under the heading: “Me want respect”. These words do not appear in this order in the text. Creative? Or not?