Personalize Claude

I’m a big fan of large language models (LLMs): both for their utility in automating some of the more tedious parts of my work, but also I’m just fascinated with them as a novel technology. If you haven’t had a chance to use them, I highly recommend you experiment with them. At the moment, the consensus killerapp of LLMs is programming. Assuming that you are coding in a decently-popular programming language, the LLMs are incredible at creating the skeleton of a working program. And if the program you are asking for is simple enough, they often nail it on the first try. They do have limitations, however. LLMs are limited in their “context window” (how much the LLM can keep in its head at once), so it won’t be able to handle really long scripts all at once. For similar reasons, working with multiple script simultaneously is also a challenge. But they are still indispensable.

At this point, there are many different LLMs available for use. The most popular one (and the one that you’ve probably have heard of) is ChatGPT. But the best one—honestly, by far—is Claude. Claude is the signature LLM of Anthropic, an AI company that started due to an internal schism at OpenAI. I can’t quite articulate what makes Claude feel like a notch above the other LLMs. But if you know, you know.

Earlier today, Anthropic announced a new feature: Personalize Claude.

Personalize Claude chats

With styles, you can customize Claude’s responses to match your communication preferences, tone, and structure. Whether you’re a developer writing technical documentation, a marketer crafting specific brand guidelines, or a product team planning extensive project requirements, Claude can adapt to your preferred way of writing.

This sounds really promising! If Personalize Claude works, then I could feed it the contents of this blog and use Claude!Eric to flood the internet with obscure probability theory content. Then Human!Eric would have more time to watch Nuzlocke videos.

This is what the prompt screen for Claude looks like:

If we click on the text that says “Choose style” next to the quill pen, then we get this menu.

It lists different possible styles for Claude. I haven’t personally experimented with any of the different styles, so I can’t comment on their various strengths and weaknesses. I have never felt any need to because the “Normal” version of Claude is quite powerful and quite flexible—given a little poking and prodding and prompt-engineering. I am also acquainted with the “Concise” version of Claude, but not by choice. Whenever site traffic is particularly heavy (like mid-mornings and early evenings), Claude will be overloaded by all the people trying to use it. As a way to manage bandwidth, Claude gets switched to “Concise” mode—which, as you would expect based on its name, is the strong and silent type. It honestly works fine; I don’t know why I am complaining about it.

But what we’re interested in aren’t the default styles, but creating our own personal style. If we click through, we are given the option to upload writing samples.

I uploaded the last couple blog posts from Critical Points.

Claude will take about half a minute to analyze your style. It will then describe the provided writing style, and also give you the option of generating various examples written in that style.

That’s a flattering description of this blog. Thanks Claude!

I decided to have Claude generate an example of “Educational Content” since that’s the closest match to content of this blog. Claude then attempted to write an explainer of Rayleigh scattering in my writing style. How did it do?

My personal verdict: Claude did terribly. This example failed both (a) as an emulation of my writing style and (b) as an engaging piece of writing in its own right. I only made it two sentences in before my eyes glazed over. I still haven’t read the entire writing sample (and I never will).

I didn’t think super hard beforehand about what my expectations were for Personalize Claude, but I can’t help but notice that I am disappointed. I definitely expected better than this. I’m not sure how Anthropic technically implemented the Personalize Claude, but based on the results, I suspect that the verbal description of the writing style that they provide is all that Claude has to work with as well.

In Claude’s defense: the idiosyncracies of personal writing styles are really hard to put into words. Humans would struggle with this task too. This actually seems like it would a fun game of telephone to be played in a Creative Writing class. You choose an obscure author that only you are a fan of. You then have to describe their writing style in words such that a classmate can write a story in emulation of their style. (Okay, maybe that was more interesting in my head.)

One fascinating aspect of LLMs is that, because they are trained on human-generated text, they exhibit a “folk psychology” similar to that of humans. In psychology, there is a concept called prototypes. Humans tend to think in categories, with certain members serving as more central examples than others. For example, an eagle would be a prototype for the bird category, while an ostrich would not. While LLMs may struggle with unique, idiosyncratic elements, they excel at capturing the stereotypical features of any genre. But who really asked for more cliche writing?