[spdustin]

Curious about the “Write in your voice” feature. Is it mostly derived from LLM stylometry, or do you blend statistical approaches like those used by spaCy? Do you capture all of lexical, syntactical, rhetorical, and semantic cues? I’d love to chat about your approach, if you’re open to any opportunities for comparing notes!

[mayop100]

Right now we use an LLM to extract a textual description of your writing style from past emails, and then we use that in the prompt.

The nice thing about this is that it's editable, so the user can customize the style to be the style they want, not necessarily exactly the style they have.

We're also investigating doing per-user model training for more refined voice... not launched yet though (and comes with tradeoffs).

Happy to swap notes -- email me: [e-mail]