Harold Pollack popularized the idea that “everything” you need to know about investing fits on a single notecard. It isn’t truly “everything,” but if you follow his 10 rules, you’ll do better than 95% of the population. The same can be done for nutrition, and the same can be done for working with AI. This isn’t your grandma’s full page of “best prompts” but some simple advice that anyone can master.
This guide isn’t for AI engineers, but for the rest of us. Scroll down to the end to get the notecard, or read on for an explanation.
Give it context
If you ask an AI to “Write a note about Rodney’s fifth anniversary at the company.” you are guaranteed to get AI slop. Fundamentally, this is because without context, the AI will default to the most common or average version of a response. In effect, you have asked it to produce boring and formulaic output.
The standard advice is to be specific and give detail about what you want, and this is good advice. “It is my employee Rodney’s five year anniversary at Microsoft, and I want to write a nice note. Include a mention of his project Contosk and his energetic attitude. It should be 5-7 sentences.” This could use even more specifics, but look how much better the result is:
The other way to bring in context is to leverage the chat history and memory. I recommend rarely starting new chats. Instead, keep one chat per topic, like “programming” or “cooking.” Last year’s models couldn’t handle long conversations, and the user experience in most chat products encourages new chats as a holdover. Don’t fall for it.
Have a conversation
I see a lot of users send one prompt to Copilot, get an answer, and then go away. Do not treat a modern AI like it is Google. You can get an answer this way that will sometimes be enough, but you are missing out on most of the value of the product. You will realize the first answer was not quite right, and you can add the missing context and get the correct answer.
The miracle of the chat interface is that it works in the same way that humans do, through language and conversation. We are naturally wired this way, and finally computers have caught up to us! Unlearn that you need special syntax to work with a computer; it is no longer true. Just talk to it!
I do not believe users need to learn prompt engineering, and I do not think you should usually use prompt engineering. Just focus on good communication techniques, and you’ll have 95% of the value.
Start over if it gets stuck
Although keeping a session going helps bring in more context and get better answers, there’s a downside as well. Think of the conversation like a journey with a friend. You start out in a direction and keep walking that way. Initially, it is easy to turn around and go another direction. But after you’ve been miles, turning has relatively little effect. Even if you start going a new direction, you remain miles down a certain path.
As you get up against the limits of what today’s AI can do, you’ll find mistakes. Go ahead and correct the AI once. But if it doesn’t take, the conversation is done. You are too far down the wrong path.
Some products will let you go back to a previous conversation turn and start a new path. Otherwise, copy what you need and paste it into a new session.
Recognize risk of hallucination
First, let me say that hallucination is much less of a problem than it used to be. Particularly in Copilot, where we spend a lot of effort on mitigating hallucination through search grounding. ChatGPT when it is using search is quite good also.
But hallucination is a side effect of how LLMs work. You could also say that every output is a hallucination – a hallucination that usually happens to be true. Key to remember is that the next word is based on probabilities. You can watch out for situations where there are lots of probable next words.
Look for situations where there are many equally likely options. For example, the date of a given event is just as likely “January 3rd” as 364 other days. Something could take place in Yakima, WA just as likely as thousands of other places. Also, consider whether what you are asking about is likely on the edge of the model’s internal knowledge. Even tiny models will give you a great overview of Newton’s laws, but be suspicious when you’re asking about the minutia of an obscure book.
With some practice, you will be able to almost always recognize when the model has hallucinated an answer for you.
Experiment and find huge acceleration
The last advice to put on the notecard is to have fun! All of these tips will become second nature through practice. My previous post gives suggestions on how to go about practicing. If a task isn’t working out, try a different part of the job.
Look for those scenarios where an AI can save 99% of your time or enable you to do 100x more. Write a web application. Personalize invites to your baby shower. Apply for a grant. Become an expert on the meaning of life.
Here are some less common tasks and patterns to try:
- Act as a specific person reviewing your document, like your boss (give context of what they are like) for a document they asked for.
- Swap roles, and have it ask you the questions leading towards working out the answer, such as a strategy for your product.
- Programming. If AI had stopped progressing last year, the gains in programming alone would change the world.
- Give you 20 ideas for the next sentence to write or direction to move the story.
But don’t take these as a guide to prompting. As Sam Altman said in 2022, “you can just…do stuff. it isn’t more complicated.”
The unprompting notecard
If you read the details above, you may not even need this notecard. It is pretty simple stuff! Someone online recently wrote, “Just talk to it, bro!”
On the other hand, maybe you’d like a reminder, or you are helping to train others use AI. Here’s a PDF of the front and back of a 4″x6″ notecard with my advice.
