Entropy AI.

Entropy started as a joke: what if a polished AI assistant was competent enough to answer you, but visibly annoyed that you asked? The project only works if the joke sits on top of an application that still feels like a real chat product.

That meant the personality could not just be a different coat of paint. The model selector, response timing, system prompts, persistent conversation state, loading behavior, and interface all had to reinforce the same idea.

The application is built with Next.js and React. Chat state is managed with Zustand and persisted locally so the user can keep a conversation without creating an account. A Next.js API route handles requests to Anthropic and selects a system prompt based on the Entropy personality the user picked.

The three modes deliberately behave differently. Standard is useful with attitude, Haiku constrains responses to the bit, and Turbo adds an intentionally ridiculous delay before returning a response. They all use the same underlying API path, which keeps the product behavior separate from the model provider integration.

Comedy is surprisingly easy to overdo in an AI product. The responses still need to be useful often enough that the app feels intentional rather than like a prompt that only insults the user.

The personality rules also have to survive real user input. Keeping the different modes distinct without duplicating the entire API implementation was the main structural concern.