Anna Bernstein is a prompt engineer at Copy.ai.
Courtesy of Anna Bernstein
Anna Bernstein is a prompt engineer at Copy.ai, which makes AI tools to generate blog posts and emails.
Her job is to write prompts to train the bot to generate high quality, accurate writing.
Here are three tips on how to write prompts to get the best outcomes from AI.
This as-told-to essay is based on conversations with Anna Berstein, a 29-year-old prompt engineer at the generative AI firm Copy.ai based in New York. The following has been edited for length and clarity.
When I was a freelance writer and historical research assistant, I spent a lot of my time scrolling through microfiche in libraries. Now I’m a prompt engineer helping to optimize the most cutting edge technology in the world.
My journey into prompt engineering began in the summer of 2021, when I met a guy at a jazz bar who, at the time, worked for Copy.ai, which makes an AI tool that can generate copy for blogs, sales emails, and social media posts.
He mentioned that Copy.ai — run on OpenAI’s GPT-3 language model — was having some trouble with the quality of its outputs and asked if I wanted to take a stab at being a prompt person. I didn’t like the stress of freelancing — plus it seemed fascinating — so I said yes, even though I was an English major and had zero background in tech.
Soon after, I got offered a one-month contract to work on executing different types of tone. At first, I barely knew what I was doing. But then the founder explained that prompting is kind of like writing a spell: If you say the spell slightly wrong, a slightly wrong thing could happen — and vice versa. Taking his advice, I managed to come up with a solution for better tone adherence, which led to a full-time job offer at the company.
Since then, the scope of my job has grown. I now help improve existing tools and create new ones with the goal of getting the AI to spit out the best responses for users.
In practice, I spend my days writing text-based prompts — which I can’t reveal due to my NDA — that I feed into the backend of the AI tools so they can do things like generate a blog post that is high quality, grammatically correct, and accurate.
I do this by designing the text around a user’s request. In very simplified terms, a user types something like “Write a product description about a pair of sneakers” which I receive on the backend. It’s my job, then, to write prompts that can get that query to generate the best output through:
Instruction, or “Write a product description about this”Example-following, or “Here are some good product descriptions, write one like this about this”
In addition to the pure prompt-engineering part of my job, I also advise on how the models behave, why they might behave the way they do, which model to use, whether we can make a specific tool or not, and what approach we should take to do that.
I love the “mad scientist” part of the job where I’m able to come up with a really dumb idea for a prompt, and see it actually work. As a poet, the role also feeds into my obsessive nature with approaching language. It’s a really strange intersection of my literary background and analytical thinking.
The job, however, is unpredictable. New language models come out all the time, which means I’m always having to readjust my prompts. The work itself can be tedious. There are days when I’m obsessively changing and testing a single prompt for hours — sometimes even weeks on end — just so I can get them to work.
At the same time, its exciting to not know what’s coming next.
Aside from people at parties not understanding my job, one of the big misconceptions I’ve noticed around AI is the idea that it is sentient, when it’s not. When it tries to talk about being an AI, we freak out, because we see so many of our fears reflected in what it’s saying. But that’s because it’s trained on our fears informed by scary, sci-fi like depictions of AI.
While writing good prompts is easy to pick up, it’s actually difficult to master. Getting the AI to do what you want it to do takes trial and error, and with time, I’ve picked up weird strategies along the way; some of my prompts are really wild in structure.
Here are some tips that can help you develop better prompts:
1. Utilize a thesaurus. Don’t give up on a concept just because your first wording didn’t get the result you want. Oftentimes, the right word or phrasing can unlock what you’re doing.
2. Pay attention to your verbs. If you want the AI to fully understand your request, make sure your prompt includes a verb that clearly expresses your intent. For instance, “Rewrite this to be shorter” is more powerful than “Condense this.”
3. ChatGPT is great at intent, so use that. Introduce what you’re trying to do clearly from the beginning, and play around with wording, tense, and approach. You can try “Today, we’re going to write an XYZ” or “We’re trying to write an XYZ and we’d like your input.” Putting an umbrella of intent over what you’re doing is always useful, and playing around with different ways to do that can make a big difference.