The home screen should answer “where do I want to continue?” rather than “which AI feature do I configure?”. Model details belong in setup and advanced settings.
Users should understand which data remains on-device and what leaves the device when a remote model is used. Privacy is product copy and behavior, not only encryption code.
The default path exposes character, conversation, attachment, send, stop, retry, and search. Samplers, prompt diagnostics, world-book rules, and provider-specific fields stay one level deeper.
Delete, overwrite, import replacement, and configuration changes must explain scope and require confirmation. Model text never bypasses local permission checks.
Draft preservation, retry, export, migration safety, and readable errors have higher priority than feeds, stickers, or decorative Agents.
Use “正在连接”, “已停止”, “没有收到文本”, and actionable errors. Do not simulate certainty, emotions, or progress the app cannot verify.
Support open character and chat formats where practical, but build an original Android information architecture and visual system.