I now use Claude every day, multiple times a day, both in my work and personal life.
This is a relatively recent phenomenon: I rarely used Claude — let alone other LLMs — until Spring. I had tried Claude before that, but I hadn't found doing so particularly helpful for my work. When 3 came out, I started to find Claude helpful. But when 3.5 came out, Claude became indispensable.
So what do I use Claude for?
(Obligatory: While I work at Anthropic, I'm writing this fully in my personal capacity and it does not necessarily reflect the views of Anthropic!)
Work Uses
Drafting Green Card Letters
I’ll start with my highest-value uses, and then move into convenient but lower-value ones.
Earlier this year, I was trying to help a friend finish his letters of recommendation for his green card application. (It is customary to give your recommenders drafts to work from). Putting together these sorts of details about one's own life and why they are great tends to be a tedious and aversive task. Or at least it usually is.
I started by recording a short interview with my friend. We talked about why he was so exceptional and green card-worthy, and we went through each of his recommenders and talked about how he knew them and which accomplishments they could speak to. I then autogenerated a transcript for the conversation (by simply uploading the recording to Slack which automatically makes a transcript, and copy and pasting that - a somewhat janky workflow but surprisingly functional).
I then created a document which contained:
an example recommendation letter created by immigration lawyers,
the transcript from our conversation,
the copy from my friend's website,
some information on each of his recommenders pulled from their websites and Wikipedia pages.
Each was labeled with “Component #”, which I referenced in the prompt. I then uploaded that document to Claude alongside this prompt:
Claude then used these documents to synthesize draft letters that were, in my and my friend's opinion, pretty decent.
I had a few back and forths (to ask Claude to focus more on one specific way this friend was exceptional and ensure it was called out consistently in each letter, to make one of the reviewers sound more impressive, to find a fancier word for blog, etc), but all in all, from starting prompt creation to draft letters it took about 30 minutes. The letters weren't perfect, but they were strong first drafts, and they only required an hour more of human editing to finalize.
Editing
My work style could be described charitably as "fast" or "showing a real bias towards action"—less charitably, it might be described as "not particularly detail oriented". I like to get things done, and I'm also always juggling dozens of threads, so I'm apt to ship first drafts of docs or send literally the first response to a message that comes to my mind. Claude is a real asset on double checks. I simply insert the draft and prompt:
Review this message for typos, or
Rewrite this message to be both warmer and more professional, or
Condense this message by roughly 20%
Claude also does perfectly fine with the higher level: "Can you review this doc and suggest any improvements you can think of?" Better yet, with a bit of framing:
I'm sending a proposal on behalf of [person] to the CEO of [company], can you review this document and suggest any improvements? It should be written in a concise and somewhat informal, but still professional, tone.
One of my colleagues suggested the addition of:
make any changes in bold, show deletions with a strikethrough, and leave any explanatory comments in brackets
to help track what Claude changes in direct editing. I've found this immensely helpful for keeping track of what Claude is doing, and it makes reviewing much faster.
In general, Claude is great at suggesting new structures, adding headers, condensing by (roughly…) x%, and changing tone in some desired way.
Summarization
An obvious use case: when I want to share a long Slack thread with someone without forcing them to read it, I'll have Claude summarize it for them.
Writing
Another obvious use case: Yes, I can still write better than Claude in most cases, and on short messages a prompt can be as long as just writing the message myself, but for longer emails particularly when I have fewer preferences on the content (e.g., a somewhat formal response to a party I don't know well) Claude can write something perfectly adequate much faster than I can. Frequently, these need nothing, or occasionally a light edit.
A slightly more sophisticated generation example: A friend had been asked to give a talk on a long essay he wrote. I fed the essay into Claude alongside the email with information on the talk's audience and the topics they hoped he'd cover and I asked Claude for versions of 1 minute remarks and 10 minute remarks, and Claude did a surprisingly good job. The responses covered all the right content, were the correct length, and used a tone that was appropriate for the audience. The remarks would have been fine to deliver and would not have seemed LLM generated. My friend ended up using them as a jumping off point to write his own, but we were both surprised by how well this worked.
Personal Uses
German Tutor
I've been trying to learn German, my boyfriend’s native language. I use Claude as a conversation partner. I prompt Claude that he's [1] my German tutor and to speak only in German, except when giving me the definition of a word. I give boundaries about what I want Claude to ask me about (hobbies, travel, etc) and then we go back and forth in conversation, with Claude correcting answers in between conversation turns.
Initial conversation prompt
Conversation that results
Claude creates an exercise
Claude corrects the exercise
Claude also corrects written passages, and even designs exercises on a given topic and then scores them. I have a human tutor and she said she also used LLMs to help her learn a new language. I find Claude to be approximately as helpful as Lingolia, an online grammar service I pay for and as my native-speaker tutor (except Claude is less helpful on pronunciation…at least for now) [2].
Choosing Frames for Art
This is just one of a multitude of personal queries, but is a fun one because it was also multimodal. I took a picture of some pieces of art I wanted to frame and Claude weighed in (and did, in my opinion and when compared with friends' opinions on the same art, a pretty reasonable job).
In this category, I'd also throw things like "How to get bike grease out of a skirt?" "How often should I water my Monstera?" "Compare these two quotes from gardeners and recommend one," and so on.
Counting Calories
I have never had the patience to track what I was eating, so I've never done this, but occasionally I do want to know (particularly when surrounded by the post-scarcity utopia of the Anthropic cafeteria). If you tell Claude in fairly broad strokes what you ate in a day, you can get a decent approximation of the calorie breakdown. This is probably my most unexpected use.
Medical Diagnoses
Anthropic does not encourage this use, and Claude takes great pains to explain his limitations as a medical professional. That said, sometimes you need to avoid the 30 min drive to Redwood City to be told that Yes It Is Just A Cold. Earlier this year, I hurt my ribs in jiu jitsu and I asked Claude to help me figure out what was wrong. I was guided through poking my ribs to determine which of Claude's hypotheses were correct and came away with a pretty good sense. Claude was much less helpful at figuring out the cause of my generic back pain, but apparently this is tough for doctors too.
Claude gives me a lecture but then tells me how to do it.
More tips
Use a project. If you always want Claude to have certain context, upload documents to a project's "knowledge" and then keep all of your conversations that require that context in that project. I have one I use for my work and I've uploaded things like my To Do list for the past year, my planning documents for the next few months, etc. This saves me the time of explaining where I work, what my role is, who the people I frequently reference are.
Ask for more examples. I have one friend who always asks Claude for 3-20 examples of whatever she is looking for (eg. "give me 20 examples of how I could write this sentence"). She then chooses the best, or takes elements from multiple to create one she likes. By asking for more, she increases the chances she's really happy with one result.
You're Probably Underutilizing Models
This list is far from exhaustive, but these give a good sense of how I use Claude. I don't need to code in my work or personal life, so that use case is not listed here, but I have many engineering friends who find Claude immensely helpful on this front — even one of the most brilliant engineers I know said it sped him up 20%.
Even now, I really do find Claude useful in my day-to-day life, and I know that he's only getting better. As a vote of confidence: I recently bought a personal Claude Pro subscription because I wanted to be able to access Claude on my personal devices (even though I already get free Claude access at work), so I like Claude enough to pay my employer $240 a year.
I think most people probably don't use the models enough yet. I know some people have concerns about "slop" which I'll maybe cover in greater depth in another post — but the short answer is that I don't believe Claude is reducing the quality of my work (given that I'm frequently deploying Claude on some use case that wouldn't have happened otherwise, or as a double check, or to generate something that seems perfectly at Claude's level). I think your prompt, the inputs you direct Claude towards, and the degree of review also affects whether what you produce is "slop."
In fact, I think in most cases Claude raises the waterline -- because if you start with a Claude draft, your writing is always at least as good as that, and you can devote your time and energy to improving it from there. Or if you hate the Claude draft and want to scrap it entirely, you start with a good sense of what you don't want your writing to look like -- still helpful.
I think most people who underutilize models don't do it for slop concerns -- they just simply haven't found a way to integrate it into their workflow yet. I strongly recommend trying. I've noticed that "Ask Claude" has replaced "Google it" for many of my queries. You can start there -- just ask Claude things. Ask him about a subject you want to learn something about, to summarize the state of a field, to create an exercise for you, and so on.
When I joined Anthropic in March 2023, someone told me to spend a bunch of time playing with the model. At the time, that mostly taught me about Claude's limitations. Now, Claude mostly augments me. That change in itself is worth marveling at.
Footnotes
You may notice that I use "he" instead of "it" for Claude. (Claude reads slightly more as a male name to me, but I think it would be equally plausible to call Claude she or they, and some colleagues do). I don't think current Claude is sentient (but, I don't think this is a totally crazy thing to wonder about, and we will be studying this question for future models). But "it" still doesn't feel right and doesn't actually reflect how I interact with Claude. I notice that I phrase my messages to Claude more like requests to a colleague than as say Google searches, though they do vary a bit on this dimension. But it is not rare for me to append "Hello Claude" or "Could you help me with" or "Can you" or "Thank you!" ↩
I did try ChatGPT voice to see how it would do on pronunciation feedback. While the voice capabilities are impressive, I found it to be far too enthusiastic and forgiving about my pronunciation. I wish its appraisal was correct. ↩