Riddle me this
Get the best from ChatGPT with these golden prompts
Afew months ago, my colleagues Cade Metz and Kevin Roose explained the inner workings of AI, including chatbots such as OpenAI’s ChatGPT, Microsoft’s Bing and Google’s Bard. Now we’re back with a new mission: to help you learn to use AI to its full potential.
People from all walks of life — students, coders, artists and accountants — are experimenting with how to use AI tools. Employers are posting jobs seeking people who are adept at using them. Pretty soon, if not already, you’ll have the chance to use AI to streamline and improve your work and personal life.
As The New York Times’ personal tech columnist, I’m here to help you figure out how to use these tools safely and responsibly to improve many parts of your life.
I’m going to spend today’s newsletter talking about two general approaches that will be useful in a number of situations.
A few common-sense warnings to start.
— If you’re concerned about privacy, leave out personal details like your name and where you work. The tech companies say your data is used to train their systems, which means other people could conceivably see your information.
— Don’t share confidential data. Your employer may have specific guidelines or restrictions, but in general, entering trade secrets or sensitive information is a very bad idea.
— Hallucinations: Chatbots are powered by a technology called a large language model, or LLM, which gets its abilities by analysing enormous amounts of digital text culled from the internet. Plenty of stuff on the web is wrong, and chatbots may repeat those untruths. Sometimes, while trying to predict patterns from their vast training data, they can make things up.
The Golden PromPTs
ChatGPT, Bing and Bard are among the most popular AI chatbots. (To use ChatGPT, you’ll need to create an OpenAI account, and it requires a subscription for its most advanced version. Bing requires you to use Microsoft’s Edge web browser. Bard requires a Google account.)
Although they look simple to use — you type something in a box and get answers! — asking questions in the wrong way will produce generic, unhelpful and, sometimes, downright incorrect answers.
It turns out there’s an art to typing in the precise words and framing to generate the most helpful answers. I call these the golden prompts.
The people who are getting the most out of the chatbots have been using variants of these strategies.
AcT As If
Beginning your prompt with these magic words will instruct the bot to emulate an expert. For example, typing “Act as if you are a tutor for the SATs”, or “Act as if you are a personal trainer”, will guide the bots to model themselves around people in those professions.
These prompts provide additional context for the AI to generate its response. The AI doesn’t actually understand what it means to be a tutor or a personal trainer. Instead, the prompt is helping the AI to draw on specific statistical patterns in its training data.
A weak prompt with no guidance will generate less helpful results. If all you type is “What should I eat this week?”, the chatbot will come up with a generic list of meals for a balanced diet, such as turkey stir fry with a side of colourful veggies for dinner (which, to me, sounds very “meh”).
Tell me WhAT else You need To do ThIs
There’s an art To Typing in The precise words and framing To generate The most helpful answers
To get results that are more personalised — for example, health advice for your specific body type or medical conditions — invite the bot to request more information.
In the personal trainer example, a prompt could be: “Act as if you are my personal trainer. Create a weekly workout regimen and meal plan for me. Tell me what else you need to do this.” The bot might then ask you for your age, height, weight, dietary restrictions and health goals to tailor a weeklong meal plan and fitness routine for you.
If you don’t get good answers on your first try, don’t give up right away. Better yet, in the words of Ethan Mollick, a professor at the Wharton School of the University of Pennsylvania, treat the bot as if it were a human intern: “When it makes a mistake, point that out and ask for it to do better.” Be forgiving and patient, and you’re likely to get better results.
ThreAd Your chATboT conversATIons
After you get the hang of prompts, you can make your chatbot more helpful over time. The key here is to avoid treating your chatbot as a web search and starting with a fresh query each time. Instead, keep several threads of conversations open and add to them over time.
This strategy is easiest with ChatGPT. Bing requires you to reset your conversations periodically, and Bard does not make it as easy to hop between conversation threads.