About 6 months ago, the Claude 3 models were released, causing me to shift to Anthropic from Open AI for most of my LLM usage.1 Recently, I was talking with a colleague who was curious about my LLM use cases. So I’ve kept track of my chat history over the ensuing two weeks to get a sense of how I use it. In this post, I detail how I’ve used Claude for pretty much everything in my life. Note: this post is going to potentially sound over-the-top in terms of heaping praise on Anthropic. Know that I am in no way being compensated for any of this.
Claude Features I like
My two favorite Claude features are easily its “artifacts” and “projects.” I haven’t seen similar features from competitors’ products, so these are a major reason for my continuing loyalty to Anthropic.
Artifacts
Artifacts allow you to easily iterate on versions of a desired output. For example, I can ask Claude for code to accomplish a task, but it almost never gets it correct on the first try. Instead, I would need to test the code it first gave me, pass it back any error messages, and iterate. For most tasks, I’ve found that it takes 8-12 iterations to finally get what I’m looking for (though this depends greatly on the complexity of the task).
With artifacts, it’s easy to toggle between each of the 8-12 versions of the desired product, as well as ask it to go back to version 7 and then start a new branch from there, etc. The artifact chain shows up in a separate side panel, so it is separate from the prompt chain, which is super helpful because it can be distracting to sift back through a prompt chain that is mixed with output.
Projects
Claude also recently released Projects. These are basically folders for prompts. This is useful for two reasons. First, it’s helpful to have a prompt history organized by topic. Second (and most usefully), Projects allow the user to share a common “knowledge base” or context across all prompts in the folder. This reduces duplication of prompt context.
Academic Research
My primary job is as an academic researcher, so this is where I get probably the most mileage out of Claude.
Coding
I’ve definitely gotten a lot of mileage out of Claude for coding tasks. These include the following:
Stata programming. Despite using Stata for over 15 years, I never really learned how to write my own programs (these would be called “functions” in most programming languages). Claude has been super useful in producing bespoke Stata programs that do exactly what I want them to. Mainly, this is useful for automating LaTeX output of results.
Unit tests (“please propose some unit tests for this Julia code and also tack them on to the end of the script in an artifact.”)
Applying new packages. Often, there’s a start-up cost when utilizing a new package in whatever software I’m using. With Claude, I can pass it the documentation of the package, tell it what I want to do, and then it can give me working code pretty quickly that does exactly what I want.
Refactoring crappily-written code (“I need your help refactoring this R code so that it can be done in a loop and possibly putting repetitive steps inside of a function”)
Writing code to fetch data from an online source and clean it.
Data Analysis
Claude is useful for all steps of data analysis, from data input and output:
How do I import delimited a tsv file in Stata?
to refining analysis output:
In Stata's estout command, how can I add "mean of dep. var" as an auxiliary stat at the bottom of the table alongside N and R2, etc?
and everything in-between:
In Stata 18 how could one estimate a regression on linked employer-employee data? I'm thinking of doing a "firm-worker fixed effects" model similar to the canonical Abowd Kramarz Margolis (AKM) model
Exploring data sources
Claude knows a lot about data sources, and it’s trivial to check whether it’s hallucinating or not. e.g.
What sorts of data about zip codes are available publicly in the US?
Managing bibliographies
Here’s my go-to prompt for create BibTeX entries from free-form text (either copied from the journal website or from the title page of the article document itself; sometimes I will even give it a screenshot of a page of the document that has the relevant bibliographic info):
I'd like you to act as a "bibtex entry maker" where I can pass you jumbled article metadata and you can provide in an artifact the BibTeX entry of the corresponding work. Please include full names as much as possible, put periods after middle initials, and Chicago-style proper capitalize the title. Please include doi (or url if doi is unavailable). If the doi is available don't put in the url. Also don't put the publication month. Also for the bibkey, make it all lower case.
This works like a charm.
Brainstorming research ideas/model mechanisms
Help me brainstorm different factors that would impact college major choices by university students in the U.S. Look through the economics of education literature. Give me the top ten reasons.
… [after prompt is answered] …
Any other reasons? More than ten?
… [after prompt is answered] …
Now repeat this exercise but instead of looking at college majors, instead consider factors that would influence interstate migration. Give me as many factors as you can that would determine whether a person moves across state lines.
Feedback on writing
e.g.
Consider the following draft of my paper introduction. Please evaluate how well it communicates the supporting ideas found in the rest of the text. Point out areas where its clarity and conciseness could be improved.
Typesetting
Claude knows LaTeX very well, so it can usually pretty easily parse LaTeX compilation errors or tie specific commands to packages, e.g.
Which LaTeX package uses $\checkmark$?
Fact-checking
Given a paragraph that I’ve written, Claude is quite good at weighing in on its factual accuracy:
Assess the validity of the following paragraph.
Explaining things
e.g.
Can you help me understand how a high-PUFA diet incites lower cholesterol levels in the blood?
Academic Instruction
Claude has also been useful to me in my role as course instructor.
Projects for each course
I find it super useful to have a separate “project” for each course I’m teaching. I can upload the syllabus and any other general information for the course as the “project knowledge” and then have a dedicated prompt history relating to the course.
This is especially useful for new course preps, where I have a rough idea of what I want to cover but don’t have a formal plan. Claude is great at brainstorming topics, making sure coverage is complete/coherent, and helping outline lesson plans and assignments.
Mentored research
I was recently helping a student come up with ideas for a directed research project and gave it this prompt, which turned out to be super useful:
I'm an econ professor advising an undergraduate who wants to do a data analysis project using stock / financial market data. What ideas do you have for how to structure the project, topics to pursue, etc? The student wants to learn how to use R to analyze stock market type of data.
In another instance, a different student had questions about a project on emergency teacher certifications. This prompt was useful:
I'm trying to figure out what economic forces are resulting in the increased prevalence of emergency certified teachers, particularly in Oklahoma.
Why would labor supply shift in the teacher labor market?
Why would labor demand shift?
What dynamics would give rise to emerg. cert.?
Is the market segmented and how would I depict this in graphical form?
Updating course content quickly
This morning, I wanted to update my PhD econometrics course slides with a new paper. Claude was able to easily figure out how to summarize the paper in a way that made sense with respect to the content on the other slides.
Administrative duties
Of course, as a professor I also have a fair number of administrative duties.
Unpacking ambiguous email chains
Sometimes I get put on an email chain where it’s not clear how I’m supposed to respond or how I’m supposed to contribute. Claude does a nice job of helping me figure this out without taxing my brain very much.
Digesting long, information-dense email messages
This is another useful application of the “summarization superpower.”
Crafting emails
Sometimes I lack the willpower to begin writing an email. Claude is helpful here in getting the gears turning.
Refining peer review reports
Claude is also useful in checking my work as a peer reviewer to see if I haven’t forgotten anything or missed any large issue with a manuscript I’m reviewing. However, I’m careful to review journal policies before using Claude since some journals explicitly have a policy that manuscripts can’t be shared with Generative AI resources.2 (Although I suspect before too long it will be easy to run high-quality LLMs on your local machine, at which point this issue will become moot.)
I don’t use Claude to write my report; rather, I use it to summarize poorly written or difficult passages in the manuscript as well as to try to find common themes in the issues I’ve identified with the paper.
Health & Wellness
I also use Claude for brainstorming about my health and wellness. For example:
“Is it ok to eat the tail of a fried shrimp?”
Analyzing bloodwork results
Considering potential pharmaceutical or non-pharmaceutical remedies for whatever happens to be concerning me (I’m, ahem, approaching a certain age and my body is not behaving like it did in my 20s)
“If I have iliotibial band pain when running, is it bad to run through the pain?”
“Can you help me come up with a physical training program that I will adhere to? I’d love to base it on kneesovertoesguy. I want it to include dips, pull-ups/low-rows, push ups, tibialis raises, calf raises, L-sit holds, seated good mornings, and elbow/wrist actions. What do you got for me?”
Religious study
I use Claude nearly every day to summarize commentary on religious texts such as the Bible or Book of Mormon. I usually am not interested in the nitty gritty details of these commentaries, but rather want to get the general idea quickly. Claude is perfect at that.
I also recently volunteered to substitute teach an early-morning seminary course for a few days. Claude was super helpful in conforming suggested lesson plans into a format that was familiar to me, as well as coming up with ideas for making the content more interesting and meaningful to the students.
Tech Support
I recently had an issue with my laptop dock in my office not properly exporting my laptop’s display to my external monitors. Within 30 minutes iterating with Claude, I was able to completely resolve the problem while also knowing the cause so that I could be on the lookout in the future.
Another recent issue was that my Fitbit watch randomly stopped syncing with my phone. I was eventually able to resolve the problem thanks to Claude’s feedback, though it took a lot longer than when my laptop dock wasn’t working properly.
General Knowledge
Claude is also useful for general knowledge since it has Wikipedia memorized. Some questions I’ve asked it:
How does Australia compare in size to the US mainland?
Which US state has the highest share of white people?
Is there a Binghamton university in the UK as well as the US?
Does Lucius Malfoy die in any of the Harry Potter books?
In Harry Potter what’s the difference between sectumsempra and rictumsempra?
What type of periodical is the Harvard Gazette?
How many cups is 150g of blueberries?
Why are strike-breakers called scabs?
What is an insolent stare? Explain as if to an 8 year old boy
Who said “now is the time for all good men to come to the aid of their party”?
Consider the chemistry of a bath bomb bubble bath accessory. Are the bubbles created from a reaction with the bomb and water?
Where is the attached photo taken from?
Conclusion
If you’ve made it this far, hopefully I’ve given you some ideas for how to better utilize LLM resources to make your life more abundant. I’ve definitely seen the quality level of my output increase dramatically these past almost-two-years, even if my quantity has stayed the same.
At the time, Claude 3 Opus came out ahead of GPT-4 for most tasks. Then, about three months later, Claude 3.5-Sonnet was released and showed even better performance. Only recently did Open AI release its “o1” model which allows for reasoning capabilities. I expect Anthropic to respond with something slightly better before February 2025.
I love that post. Thanks, Tyler.