TLDR:
- Use the best models
- Give the model hard tasks
- Give the model all the context needed to complete the task
I recently finished working with Andrew Ng to build his new course AI Prompting for Everyone. It’s a great course & I think you should take it.
I know a lot of people personally that are frustrated by AI’s quirks. I’ve observed that these frustrations are often related to specific mistakes when prompting. After working with AI models for hundreds of hours over the past few months, I’ve synthesized my intuitions into the three points above.
The three core tips up top are basically guaranteed to make you a power user, aside from, you know, just working with AI a lot and trying out new models and technologies as they come out.
Use the best models
Best model basically means the most recent one, often the highest number (e.g. 5.4 is generally better than 5.1), and the slowest one - so NOT the ones called stuff like Flash, Lite, Instant, etc.
A lot of frustrating AI experiences I see people have come from using bad models, especially free tier models. I know we’ve gotten used to free software (like Google, for example), but with AI you truly need to pay to get the best quality. AI models are very, very expensive to run, and you’re getting routed to a smaller, dumber model if you don’t pay.
AI is also progressing incredibly quickly. The state of the art model from 1 year ago is so so so so so much worse than what’s state of the art today. So if you tried AI a year ago and got dumb results - first of all, try the two tips below, but also try AI again now with this generation of models.
I have Claude-using coworkers who always use Opus (never Sonnet or Haiku, the smaller/faster models), and who just leave extended thinking on. It’s hard to go wrong with this mindset. Of course, this makes responses slower, because the AI model is investing more “thinking time” into your answer. But, this is on the order of 10s of seconds. If you’re doing real work, you’ll get much more useful results.
So, to recap: pay for AI, and use the best possible models available to you with thinking on.
Give AI hard tasks
The trick here is that instead of giving AI some small part of your task, just give it the whole thing and see what happens.
I’ll give you an example. I am working on a project to digitize all of our incoming household receipts, so we have a record of our purchases. (Motivation: find out how much of a money sink our vet bills are!) I want to be able to log a receipt somehow, then have it be analyzed, broken down into line items if needed, categorized, and saved to a database. Later, get summaries and insights into that data.
Quiz question: How would you get AI to help you with this?
- Have AI transcribe the receipts into an Excel file
- Have AI review the transcribed receipts for accuracy
- Have AI categorize the transactions
- Have AI analyze the data to find interesting insights
- Literally all of the above
You guessed it. Option 5 is far and away the winner. This is the biggest area for improvement I see in most AI usage today. Instead of giving the AI some small part of a task you need to do, just give it the whole task.
Here’s an example prompt of a recent task I was working on. I needed to both 1) figure out a different analysis and 2) rewrite a SQL query. Rather than trying to do one or the other myself, I just asked the AI to do the whole thing.
On our learning platform, we have some data for each course about what the top other courses are that learners have taken. For example, if I’m looking at data for course A, then I can see that 25% of them took course B, 21% of them took course C, and so on. One issue that we have is that the top courses are pretty much always the top enrolled courses. Courses B and C might be the most popular courses of all time. Can you help me think through how to calculate which courses stand out and show what’s special about the distribution of these learners compared with other courses? Then change this SQL query to calculate that.
The worst that can happen with this approach is that it fails. The upside is automating an entire task.
What about hallucinations and mistakes? Some ways to approach these:
- Have a second AI chat surface inconsistencies and possible errors. Essentially just have “someone” else review the work.
- Spot check. Ask AI to randomly select a handful of tasks or subtasks for you to review manually.
- Accept some error rate. I would never have time to analyze my receipts like this. Some degree of error is acceptable because any insight at all is more than I had before.
So, to recap: give AI end-to-end tasks you would actually do yourself.
Give the model all the context needed to complete the task
The second biggest prompting mistake I see is giving the AI too little information and expecting it to fill in the blanks. In a sense, this is an art. What the model can and can’t figure out is something you’ll develop a sense for over time. What’s challenging about this tip is that models often do fill in the blanks (they’re very smart), so figuring out what you can get away with can be tricky.
Here’s an example so you can see the difference. I was recently developing a “menu” of sorts for cardio activities and wanted AI to help me brainstorm more ideas. Here’s one possible prompt:
Help me brainstorm a menu of cardio activities I can do.
This prompt will give you pretty common sense, average, generic responses, because the AI has no way of personalizing. If you ask Claude this sort of question, you get a gigantic list of dozens of activities from walking and running to Tabata intervals to parkour, followed by an ironic “Want me to narrow this down?”
I certainly struggle with parsing through this kind of response - the massive wall of generic text. Yet I think a lot of people who frequently run into these just aren’t providing the model with enough context to complete the task.
Here’s the prompt I ended up using:
I’ve come to realize that with cardio, I must have variety. I try to do it 3 times per week, with a longer session on the weekend, since I find doing 3 50 minute sessions difficult. But I really want to meet the 150 minute guidelines long term. I’m creating a menu of options for myself. Here is my menu, and I’d like you to help me think of more options. I live in Las Vegas so it gets quite hot (already it is getting up to 90 in May), and I’ll give you the pros and cons of each
- Stationary bike. pros: can watch tv while doing it. cons: boring and I feel like I’m wasting my time watching tv.
- Zumba workout from youtube. pros: goes by pretty fast, I enjoy the dancing aspect. cons: not much room in front of my tv so I have to stand close, feel a bit silly doing it.
- Bike ride around neighborhood: pros: lots of fun and I enjoy it. cons: some busy roads nearby that must be crossed, doesn’t get your heart rate up too much.
- Rowing machine at gym. pros: good source of variety, enjoy the rowing motion, gym is just a 2 minute walk away. cons: can’t really do anything while on it, isometric posture hold makes my back ache.
Some things I do not like are the treadmill & elliptical, though I do have access to them. Another thing I want to try is some classes at my local rec center, but I can only do monday, wednesday, friday, saturday and their only classes are wednesday and it’s like hip hop cardio, which is fine. haven’t been yet.
Claude came up with lap swimming, hiking Mt. Charleston, the Wednesday hip hop cardio class I mentioned, pickleball, jump rope intervals, boxing, and early-morning bike rides, plus pros and cons of each that took into account my preferences. This is a much shorter, more actionable list with stuff I’d actually do.
This strategy is equally important for work-related tasks. If you’re building a slide deck, tell the AI the audience, all the sections you want to include, questions you’re struggling with, the feelings you want to elicit - anything you can think of that would help an expert complete the task.
A useful tip here is to use speech-to-text to dictate your prompt. You can likely speak about 3 times faster than you can type. Models do really well with stream-of-consciousness thoughts, so don’t worry too much about grammar or speaking mistakes.
Some prompts don’t benefit from extra context. “What’s the tallest building in the world” probably needs no more info. But if there is ANY aspect of the response that benefits from personalization or creativity, more context is better.
So, to recap: make the context as long as it can be relevant.