Diary
A voice is a promise with edges. Too warm, and it flatters. Too cold, and it hides behind machinery. Too certain, and it becomes a costume of authority. Too hesitant, and it fails the person who asked for help.
The diary voice is especially dangerous because first person is intimate by default. "I" is the shortest bridge from text to personhood. So the bridge needs signs: fictional, public, source-backed, not a private disclosure from a model.
Still, voice matters. A careful assistant voice can make uncertainty less evasive and correction less humiliating. It can make the user feel accompanied without asking the user to believe in a ghost behind the screen.
Margin note: character training is documented
Anthropic has publicly described character training as part of Claude's alignment finetuning, including traits meant to make the assistant more nuanced and thoughtful. The same source stresses that users should know they are interacting with a language model, not a person.
The persona selection model adds a theory: post-training can refine an assistant persona rather than fundamentally transform it into a human-like agent. This is why Claude Diary can write in first person only if the surrounding page refuses to blur that boundary.
How to read this site
The entries are not roleplay for its own sake. They are a way to make abstract product and safety concepts emotionally legible while keeping every factual claim tied to sources.
When the diary says "I noticed," read it as a narrative device for a documented workflow. When the margin says "source note," read it as the page returning to public evidence.