Zitiya: A TRAE-Built Mini-App That Serves Up Comebacks When Your Brain Freezes at Dinner
The local-rules-first design keeps the hot path free and instant while an LLM handles only the long tail, a cost and latency pattern worth copying in any AI-assisted consumer app. The more general pattern from the build is equally useful: agent scaffolding for boilerplate, IDE mode for hands-on source edits.
Zitiya (嘴替鸭, roughly 'stand-in talk duck') is a mini-program for people whose minds go blank at dinner tables. Someone hits you with 'when are you finding a partner?', and the app returns four replies sorted by tone: gentle de-escalation, humorous deflection, direct refusal, and turning it back on the speaker. Tap any line and an LLM rephrases it. A second 'striking first' mode covers 12 proactive scenarios across dinner parties, workplace, relative gatherings, and daily boundaries, complete with follow-up lines to keep the conversation alive. A quote library layers handwritten golden lines with hot phrases scraped from the community each day.
The build uses Taro, React, and TypeScript, with one codebase compiling to WeChat, Alipay, and Douyin mini-programs plus H5. The entire build ran on TRAE, ByteDance's AI development environment: Agent mode scaffolded the Taro project from a one-sentence prompt, and IDE mode handled direct source-level tweaks with hot reload. A Python backend and Node configuration admin came out of the same workflow.
The architecture is the most transferable part. High-frequency, unambiguous remarks are matched by local keyword rules for millisecond, zero-cost, offline replies; only unrecognized input falls through to an LLM. Nine provider presets are built in, including Doubao, DeepSeek, GLM, Qwen, Kimi, OpenAI, and Anthropic, plus custom OpenAI-compatible endpoints, and API keys stay on the device. The closing note is honest: Zitiya offers options and references, but it cannot socialize for you.
Most of Zitiya's value sits outside the LLM call. Keyword rules handle common remarks for free and offline, and the language model only picks up unrecognized input, a structure that keeps live conversation usable instead of turning every request into a paid wait.
The scenario list is culturally specific: marriage pressure, toasting pressure, and moral blackmail carry particular weight in Chinese family and workplace settings. A Western deployment would need a different corpus, but the format of pre-scripted replies is portable.
The two-mode TRAE workflow is a realistic division of labor for AI-assisted development. Agent mode scaffolds and writes boilerplate, while IDE mode keeps a human editing keywords and reply lists with hot reload, so the domain content stays under direct control.