For a brief time in 2023, prompt engineering was hailed as an exciting, high-paying career path for a select few—until it quickly became clear that, far from being a dedicated role, it was a skill everyone could (and should) incorporate into their work and lives.

That doesn’t mean every ChatGPT prompt you enter is going to return the results you want. It takes clear communication, iteration and sometimes, no small amount of patience. AI is powerful, but it’s not a mind reader. It works with what you give it.

As models have become more advanced, how we compose prompts has evolved, too. Some advice that rang true years ago is now obsolete, meaning if you’re still following it, you may not be using LLMs to their full potential .

You’ve probably come across articles advising you to ask AI to inhabit the mind of some sort of professional—”you’re the head chef at The French Laundry, explain how to sauté bell peppers,” or “you’re a Formula 1 race engineer, help me find trustworthy mechanics in my area.”

But research from Wharton says you can save the role play, since it doesn’t actually improve factual accuracy. “Despite being recommended by major AI providers, assigning domain-specific expert personas doesn’t improve performance on difficult factual questions, even when the expertise matches the question domain,” the authors write.

If you’re truly attached to this part of prompting, there is evidence to suggest that it can be helpful for providing context or influencing the tone or approach of the answer. Just don’t feel obligated to couch your questions this way, since a clear, to-the-point prompt will generally lead to the same output.

When Words Fail, Try A Photo

On a recent vacation with my family, our Airbnb’s TV quit working. I wasn’t familiar with the setup, and my kids were growing restless for our family movie night to start. This was not the time for tinkering.

Instead, I opened ChatGPT, described the situation, and uploaded a photo of the TV’s make and model number. Almost immediately, it was able to troubleshoot the issue, giving me step-by-step directions for what happened and how to fix it.

They say a picture is worth 1,000 words, and when it comes to directing AI, this is definitely accurate. From diagnosing a mysteriously wilting plant to deciphering a hand-drawn diagram on a white board, today’s models work far better with a simple phone photo than a garbled description. Typing out every detail is slow and easy to get wrong. A photo literally presents the whole picture at once—plus, the model will likely pick up on details you may have inadvertently left out.

In the early days of genAI, it was common to overcommunicate: we assumed the more information it had to work with, the better the outcome. Now, we know better. Like humans, AI can get confused and lose the plot.

Researchers have referred to this as “context rot,” and it happens consistently. When Chroma, an AI research and infrastructure company, tested 18 leading models, every single one performed worse as the input grew longer, even on simple tasks where the extra material was relevant.

Now we know that the goal shouldn’t be to dump every imaginable data point into the prompt field. Instead, pare back to the information it actually needs to do its job. Say I want advice for planning my son’s birthday party. I don’t need to give ChatGPT my life story or walk it through every party we’ve ever thrown. I do tell it how old my son will be, that he’s obsessed with space and that we’re hosting eight kids in a small backyard on a Saturday afternoon.

Think of AI as a capable person, not an omniscient being that requires no context or parameters. The more you do it, the more you’ll get a sense of the sweet spot that gets you the answers you’re looking for.

Sometimes, ChatGPT’s ability to build from your previous chat history can be a hindrance. Maybe you’re working on one project one day and a completely different one the next; even though the two have nothing to do with each other, the model will often try to link them.

In this case you have a few options: if you’d prefer to stay in the same chat window, you can simply ask it to forget everything you’d previously discussed, wiping the slate clean. Another option is to use temporary chat, which functions as an incognito mode that neither relies on past conversations nor will store context for future ones.

While it may not be the lucrative career path it was once made out to be, prompting is a skill everyone can master. But models are changing quickly, and the advice doled out in 2023 isn’t helpful today. By continually brushing up, you can make sure you’re using AI as effectively as possible.