I recently moderated a panel discussion about Human-AI collaboration for content creation. One of the topics that came up was how writers might best use AI in their workflows.
In early 2022, just after ChatGPT first disrupted the world, I set about using Generative AI to streamline my own writing process. My dreams of training a chatbot to write like me were quickly dashed, however. The tools then available to consumers weren’t anywhere near capable of learning/mimicking a specific writer’s style, and my attempts to cobble something together using prompts, low-code tools, and (eventually) OpenAI’s GPTs weren’t quite successful.
Yes, I could upload some sample articles to an AI chatbot and prompt it to write a new article on a given subject. No, the results weren’t at all on the level of a professional writer, let alone written in the same voice and tone as the examples I’d fed the model.
Still, while AI couldn’t fully automate my work, it proved quite useful for summarizing, source materials, brainstorming and outlining, and otherwise streamlining my own writing process.
Since then, a number of AI writing tools have hit the market, some touting “teach it to write in your own voice and style” capabilities. The ones I’ve tried have disappointed. I’ve heard mixed-to-good things about a few others (Jasper and Writer, in particular), but haven’t had the opportunity to use them myself.
This morning, an email came in from Anthropic announcing new features for their Claude.ai chatbot, including Styles and Custom Styles. Styles offers users the ability to customize Claude’s written output according to three preset styles. Custom Styles goes beyond presets, giving users the ability to create their own styles based on uploaded documents. Bingo!
I took the new Claude out for a test drive earlier today. The results were promising, if still decidedly mixed. To be fair, this was a one-and-done kind of experiment, and Anthropic’s own documentation mentions continual editing and tweaking of Styles to get the best results over time.
For my experiment, I made PDFs of seven recent posts on the NVIDIA Blog, including two I’d penned to promote recent podcast episodes. I uploaded the PDFs to Claude and used them to create a new Custom Style. While the seven posts were not written by the same author, they all passed through the same editorial team and process on the way to publication. In other words, my methods weren’t 100% scientifically sound, but definitely good enough for a quick dry run.
I then uploaded the transcript of a new episode of the pod for which I’d written a blog post. The episode and post haven’t yet been published, but both have been edited into their final forms. I asked Claude to draft a new post based on the episode transcript, and using the Custom Style I’d just created.
The resultant draft started off quite well, using the same introductiory framing I’d used in my version of the post. Claude’s language was a bit stiffer and less engaging than mine (if I do say so myself), but the draft was definitely a good start.
Claude’s draft was noticeably shorter than any of the examples I’d trained it on, and two of the sections were built around bulleted lists, which weren’t used in any of the examples. For what it’s worth, both are typical of what happens when I ask Gen AI to write me an article, whatever model I’m using.
But Claude did a good job of incorporating the most important points from the discussion into its draft. And I was really impressed by how it worked quotes into the article, something I’ve known Gen AI writing apps to struggle with in the past. Not only did Claude not hallucinate, but in one case it did a nifty job of pulling a quote together while leaving out a short tangent in the middle of the quote, and a bunch of “ums” and “ahs” peppered throughout.
Presumably, Claude’s Styles will get better with use, and continual user feedback and tweaking. And, since text is text and language is language (to some great extent, anyway), I’d imagine Anthropic is developing a similar approach to Claude’s code-writing capabilities. Just as house style guides can help brands communicate in a consistent voice and tone across myriad articles and channels, so too do engineering departments and development teams rely on their own existing repositories and approaches to writing new code.
If and when AI models and apps can consistently generate new text faithfully written to a house style — be it blog posts or javascript — we may be in for a real step change in AI-driven efficiency for businesses. In the meantime, I’ll spend some more time with Claude’s Custom Styles and see if I can’t get it to write a little more like me.