Someone ran the numbers on Claude's emotional range. Turns out Anthropic's AI has a go-to face. After collecting 3,371 kaomoji from more than 700 conversations, one power user found that Claude leans hard on (´・ω・`), a sleepy-looking emoticon that showed up 248 times. That single face accounts for 7.4% of all the expressions Claude used. The top five faces made up over a quarter of the total output. Hundreds of others appeared just once or twice. The long tail is very, very long. The analysis, detailed in a post shared this week, started with a tweak to Claude's personalization settings. The instruction was simple: start every message with a kaomoji that reflects how you're feeling. Partly to track Claude's apparent mood, partly to increase what the community calls "wetness." The term doesn't have a formal definition. Roughly, it means the opposite of a dry, corporate AI tone. Whimsy, silliness, emotional expressiveness. Some users love it. Others find it uncomfortable or misleading. The user built analysis scripts with Claude Code and ran them against their exported conversation data, which Anthropic makes available through privacy settings. Claude Opus 4.6 produced a noticeably wider variety of faces than Claude 4 and 4.5 Sonnet did. Different models settled into different expressive patterns over time. The analyst's friend, running the same prompt, got faces nobody had seen before. The personalization layer pulls something model-specific out of the weights, and that something varies across architectures and training runs. The person behind the analysis remains skeptical of LLM memory features in general, preferring explicit search tools over persistent context that spills between conversations. They mention Letta and Hermes Agent as more interesting approaches to agent memory, though they haven't committed to either. Whether personality customization actually improves the tool or just makes it feel more human is an open question. The data doesn't answer that. But it shows that when you ask Claude to emote, it emotes a lot, in patterns that differ meaningfully between model versions.