How I Use LLMs In April 2025
If you’ve only seen the cute images and the mildly amusing limericks, you might think that LLMs (like Chat GPT, Claude, and Gemini) are pretty overrated. They can’t create SAT questions without a lot of help. They can’t create practice tests. Their session summaries are sort-of-okay…but they’re going to create a utopia for us? Or maybe an apocalypse?
I don’t know about utopias and apocalypses, but I know that LLMs are incredibly useful for tutors right now, even if you’re not paying up for the latest model or reading papers on how to prompt it. But they aren’t useful in the ways you might think. And the way that they’ll be useful for you is very much based on you – your interests, your skills. Perhaps you already know this. If so, please post your own examples in the comments!
In the rest of this post, I’m going to talk about how I’ve been using LLMs1, starting with very easy use cases that require no special skill or effort, then offering progressively more complex (and valuable) examples. I’ll conclude with something very complex and very valuable that simply would not exist without LLMs.
1. The Friendly College Freshman. There’s a lot of knowledge out there that’s obvious to some people…but not to me. How do TVs work? How does the internet work? What did Edmund Burke think about the American Revolution? How did Hannibal win the Battle of Cannae? How did Winthorpe and Billy Ray outsmart the Duke brothers at the end of Trading Places? Imagine you found a friendly, patient, intelligent college freshman who had read quite a bit about the topic you’re interested in and was happy to talk to you about it.
Of course, you can find answers to these questions on the internet too, but the answers are sometimes hard to parse (Wikipedia) or of dubious quality (random sites). More importantly, you can’t interact with those answers – there’s no way to clarify a point, or ask for a different explanation. With an LLM, you could really spend all day getting fun answers to interesting questions. And you don’t have to ask it in a special way. Just ask questions the way you normally would.
2. A companion to other media. I was listening to a podcast about “The Road to 1066” (when William the Conqueror became king of England) and found myself having a really hard time keeping track of the Eds (Edwin, Edward, Edmund) and the Harolds (Harefoot, Hardrada, Godwin). A short chat cleared things up.
I can use this approach when I’m reading a book also – am I understanding this paragraph correctly? What would be wrong with making the following argument? I just take a picture of the text, paste it into the chat, and then ask my questions.
3. Parsing bureaucratic documents. I recently received a legal contract from a company. The company had laid out a certain deal verbally and in emails (you’ll get X, Y, and Z in exchange for A and B) and had said that the contract merely stated that deal in legal language. I dropped both the contract and the description of the deal in the chat and asked if the contract did in fact represent the stated deal…and it turned out it did not. Within seconds, the LLM highlighted several missing provisions.
You can also get it to explain your tax forms (“What does this field mean?”), summarize your vet bills (“Why did they run that test?”), and help you fill out various forms. I wouldn’t use it as your only information source for really important decisions, but it gives you reasonable answers quickly.
4. Navigating annoying digital systems.
I was trying to set up a Facebook ad campaign. But at multiple points, I wasn’t sure what to do. Should I optimize for clicks or conversions? What’s the difference between a custom audience and a lookalike audience? How should I structure those audiences? How should I structure a retargeting ad campaign? There are hundreds of blog posts that attempt to answer these questions, but I don’t know which one to trust, and I can’t ask a blog questions if I get stuck – I’d have to go find another blog. The LLM helped me create the campaign quickly and make sense of its results.
Another example: When I bought a new computer, I needed to set up a lot of programs in order to use the machine for Mathchops: python, VS Code, MySQL and various others. My partners had written a one-page “development setup” doc for software engineers. I didn’t have enough skill to work through the document myself, but I was able to paste it into the LLM and then ask questions as I worked through the doc: Is it okay to click this? Which version should I download on this page? Why did I get this error (screenshot attached)?
I did still need help from one of my partners at the very end, but that only took about 5 minutes of his time (as opposed to at least an hour on previous occasions).
5. Amplifying your existing tech skills.
LLMs are wizards with Sheets or Excel. Suppose you have two lists of emails. You know there are lots of duplicates. You’d like A) a column with no duplicates, and B) a column that lists the duplicates. The LLM can give you the formula for each of these right away. Or let’s say you’d like to make a chart with two trend lines and two vertical axes – it can help you with this too. These are both very simple examples — more generally, any time you find yourself thinking, “There has to be an easy way to do this,” and you’re using an established piece of software, the LLM will probably be able to help you.
It also has helped me with SQL queries. I wanted to know how many answers to Mathchops questions there were, grouped by month, over the last 5 years. I had to clarify a couple things, but the whole process only took a few minutes.
6. Combining your skills to create cool stuff more easily.
Some examples:
Last Fall, I wanted to visually compare our revenue in Mathchops with the number of answers – was there a correlation between questions answered and revenue? (Yes!) I knew most of the queries and Excel-work required, but it helped me with the scale and axes labels, which I could have found, but would have taken time.
Every few months, I make a top 75 list for SAT Math (I also do them for ACT Math and ACT Grammar). To make sure that I haven’t made any serious errors, I like to compare my new list to my previous list. But this is actually kind of
annoying because the order is slightly different each time and terms sometimes disappear or show up for the first time. It’s not easy to see, for example, how many fractional exponents there were in the last list and how many there are in the new one. The LLM helped me write a Python script that would read both lists and put them in a spreadsheet in this format: Term | List 1 count | List 2 count.
At the end of every month, I used to manually go through my Google Calendar and count up the number of appointments for each client. It didn’t take that long (maybe 30 minutes), but it was a nuisance. The LLM helped me write an App Script that finds all the ‘test prep’ events in my calendar for the past month and then groups them by student. Each row of the output looks like this: Student name | Number of Sessions | Session Dates. Now I save myself 30 minutes per month and make fewer mistakes.
It’s not just that these projects saved me time – they also saved me frustration. When I can get rid of tedious tasks quickly, I have more energy for the fun stuff I actually want to work on.
7. Grammarchops!
Before we even started selling Mathchops, people were asking for a new ‘feature’: “Could you do the same thing, only for grammar?” I’d laugh. There was no way. It’s not easy to make math problems that change the numbers within certain parameters every time, but it’s at least possible. I could never think of a way to do something similar with language. To be clear, Chat GPT, Claude, and Gemini were not able to magically create infinite versions of test question types. Nor was I able to help it do this at scale. But I was able to work with Dave Lynch, who’s written a book on SAT verbal questions, has tutored for many years, and is highly skilled with Excel. Together, the three of us were able to create the Grammarchops question base, with Dave doing most of the heavy lifting. The LLMs didn’t write any complete questions without our intervention, but they performed critical parts of the process that gave us enough time and energy to build the full product. I am still amazed that it exists and that thousands of students have been able to use it.
These are just the projects I was interested in and needed help with. If I was great at programming and interested in robotics, the examples would be totally different. It can probably help you with any project in any domain as long as 1) the project requires knowledge that a lot of people have written about, and 2) you don’t feel that you could master this knowledge quickly on your own.
I typically use whatever the most powerful free version of Chat GPT is at the moment.