By Associated Press
By KELVIN CHAN, Associated Press
LONDON 鈥 If you鈥檙e using ChatGPT but getting mediocre results, don鈥檛 blame the chatbot. Instead, try sharpening up your prompts.
Generative AI chatbots such as OpenAI鈥檚 ChatGPT, Google鈥檚 Gemini and Anthropic鈥檚 Claude have become hugely popular and embedded into daily life for many users. They鈥檙e powerful tools that can help us with so many different tasks.
What you shouldn鈥檛 overlook, however, is that a chatbot鈥檚 output depends on what you tell it to do, and how. There鈥檚 a lot you can do to improve the prompt 鈥 also known as the request or query 鈥 that you type in.
Here are some tips for general users on how to get higher quality chatbot replies, based on tips from the AI model makers:
Be more specific in your prompt
ChatGPT can鈥檛 read your mind. You need to give it clear and explicit instructions on what you need it to do.
Unlike a standard Google search, you can鈥檛 just ask for an answer based on some keywords. And you鈥檒l need to do more than just tell it to, say, 鈥渄esign a logo鈥 because you鈥檒l end up with a generic design. Flesh it out with details on the company that the logo is for, the industry it will be used in and the design style you鈥檙e going for.
鈥淓nsure your prompts are clear, specific, and provide enough context for the model to understand what you are asking,鈥 ChatGPT maker OpenAI advises on its help page. 鈥淎void ambiguity and be as precise as possible to get accurate and relevant responses.鈥
Refine and rewrite your request
Think of using a chatbot like holding a conversation with a friend. You probably wouldn鈥檛 end your chat after the first answer. Ask follow-up questions or refine your original prompt.
OpenAI鈥檚 advice: 鈥淎djust the wording, add more context, or simplify the request as needed to improve the results.鈥
You might have to have an extended back-and-forth that elicits better output. Google advises that you鈥檒l need to try a 鈥渇ew different approaches鈥 if you don鈥檛 get what you鈥檙e looking for the first time.
鈥淔ine-tune your prompts if the results don鈥檛 meet your expectations or if you believe there鈥檚 room for improvement,鈥 Google recommends in its prompting guide for Gemini. 鈥淯se follow-up prompts and an iterative process of review and refinement to yield better results.鈥
Consider the persona and audience
When making your request, you can also ask an AI large language model to respond in a specific voice or style.
鈥淲ords like formal, informal, friendly, professional, humorous, or serious can help guide the model,鈥 OpenAI writes.
You also tell the chatbot the type of person the response is aimed at.
These parameters will help determine the chatbot鈥檚 overall approach to its answer, as well as the tone, vocabulary and level of detail.
For example, you could ask ChatGPT to describe quantum physics in the style of a distinguished professor talking to a class of graduate students. Or you could ask it to explain the same topic in the voice of a teacher talking to a group of schoolchildren.
However, there鈥檚 plenty of debate among AI experts about these methods. On one hand, they can make answers more precise and less generic. But an output that adopts an overly empathetic or authoritative tone raises concerns about the text sounding too manipulative.
Add more context and examples
Give the chatbot all the background behind the reason for your request.
Don鈥檛 just ask: 鈥淗elp me plan a weeklong trip to London.鈥
ChatGPT will respond with a generic list of London鈥檚 greatest hits: historic sites on one day, museums and famous parks on another, trendy neighborhoods and optional excursions to Windsor Castle. It鈥檚 nothing you couldn鈥檛 get from a guidebook or travel website, but just a little better organized.
But if, say, you鈥檙e a theatre-loving family, try this: 鈥淗elp me plan a weeklong trip to London in July, for a family of four. We don鈥檛 want too many historic sites, but want to see a lot of West End theatre shows. We don鈥檛 drink alcohol so we can skip pubs. Can you recommend mid-range budget hotels where we can stay and cheap places to eat for dinner?鈥
This prompt returns a more tailored and detailed answer: a list of four possible hotels within walking distance of the theater district, a seven-day itinerary with cheap or low-cost ideas for things to do during the day, suggested shows each evening, and places for an affordable family dinner.
Put limits around your request
You can tell any of the chatbots just how extensive you want the answer to be. Sometimes, less is more.
Try nudging the model to provide clear and succinct responses by imposing a limit. For example, tell the chatbot to reply with only 300 words, or to come up with five bullet points.
Want to know all that there is to know about quantum physics? ChatGPT will provide a high-level 鈥済rand tour鈥 of the topic that includes terms like wavefunctions and qubits. But ask for a 150-word explanation and you鈥檒l get an easily digestible summary about how it鈥檚 the science of the tiniest particles that also underpins a lot of modern technology like lasers and smartphones.
Is there a tech topic that you think needs explaining? Write to us at onetechtip@ap.org with your suggestions for future editions of One Tech Tip.