Trizenta
Shipped product (0→1, sole designer)已上线产品（0→1，独立设计师）
Year
2026
Duration
~7 weeks, design to live MVP约 7 周，从设计到 MVP 上线
Overview
Trizenta is a live AI travel decision engine. Two flows run on one engine: a solo search and a group coordinator, both returning three bookable trips. As the sole designer on a four-person team, I owned it end to end, from brand and IA to two product flows, and delivered the running frontend to engineering myself.Trizenta 是一个已上线的 AI 旅行决策引擎。个人搜索与多人协调两条流程共用同一套引擎，都输出三个可直接预订的行程。四人团队中唯一的设计师：品牌、信息架构、两条产品流程，以及交付给工程团队的可运行前端。
My role
Sole Designer (Product Design + Design Engineering)独立设计师（产品设计 + 设计工程）
At a Glance
An AI travel decision engine that turns a vague feeling into three bookable trips替你收敛选项，不替你做决定
Solo search returns three bookable trips, each scored by signals.个人搜索返回三个可预订行程，各自按信号评分。
What:
A short-flow decision engine, not a planner. Solo search and a group coordinator, both returning three bookable trip cards.决策引擎，不是行程规划器。流程只有一步：一次输入，三个可直接预订的方案。
Why:
The first wave of AI trip planners died by making planning longer. People want to decide, not to plan.选项从来不缺，缺的是把它们收敛成一个决定。规划类工具擅长把信息铺开，用户真正卡住的地方在之后：在几个都不错的方案之间做取舍。
How:
Restraint. AI reads the intent behind messy input, but the person makes the call. Three options, scored for instant comparison.AI 只负责读出杂乱输入背后的意图，决定权不交给它。三个选项，各自按你输入中的维度评分，可以直接比较。
Who:
People stuck in decision paralysis, and groups that can never converge on one trip.决策成本过高而迟迟定不下来的人；以及难以就同一趟旅行达成一致的团体。
Landing page落地页
Discovery
The bottleneck was never options. It was the decision.这从来不是一个信息问题
Framing
ChatGPT gives you ideas, Google gives you a thousand tabs, and the first wave of AI trip planners only made deciding take longer. But the bottleneck was never information, it was the choice itself. Talking to travelers, the same thing kept surfacing. Not “I can’t find options,” but “I’m stuck.” Someone who loves the research but has no time for it. A couple who want opposite things from the same trip.访谈里反复出现的不是「我找不到选项」，而是「我卡住了」。有人享受做功课，却挤不出整块时间；一对伴侣，想从同一趟旅行里得到相反的东西。两种情况的共同点是：他们要的不是更多选项，而是一个可以停下来的决定。
Design
AI travel decision engine个人搜索流程
Escaping the chat box
Restraint had to be built, not stated. A chat box is the zero-friction way in, but staying inside the chat paradigm would have meant endless back-and-forth, exactly the long flow we were rejecting. So the experience leaves the chat frame the moment you submit: no sidebar, no history, three trip cards take over, and the prompt collapses into a search bar at the top. AI reads the intent. The interface does the deciding work.聊天框是摩擦最低的入口，但留在聊天范式里就意味着多轮往返，而那正是这个产品要避开的形态。所以提交的那一刻，界面就离开了聊天：没有侧边栏，没有历史记录，三张行程卡片接管画面，输入框收起成顶部的搜索栏。意图由 AI 读取，决策由界面承载。
Early exploration
An early exploration kept the LLM chat structure. I dropped it: it pulled the product back into the long flow we were trying to escape.早期探索保留了 LLM 的聊天结构，后来放弃：它把产品拉回了本想避开的长流程。
Where it landed
The shipped version leaves the chat frame entirely. Cards take over, the prompt collapses to a search bar, and deciding happens in one screen.最终上线的形态：一屏之内完成决策，没有对话历史需要回溯，也没有第二轮追问。
Signal mechanism
Each card is scored by signals, dimensions pulled from your prompt, so three options compare at a glance.评分维度直接取自你的输入，因此三张卡片是按你在意的标准排的，而不是一份通用榜单。
Design
GroupTrip CoordinationGroupTrip 多人协调
Converging without the group chat
Groups don’t fail to find options. They fail to converge, and a group chat just turns it into who has the loudest voice or the most patience. GroupTrip is built so a group can land on one trip without that.团体的问题不是找不到选项，而是无法收敛。群聊会把这件事变成谁声音大、谁更有耐心。GroupTrip 要让一群人绕开这一步，直接落到同一趟行程上。
Collecting state收集阶段
Everyone drops in
You drop in a prompt, in your own words, anonymously. Beside it, a quiet bar holds everyone else’s, hover and they unfold. A 24-hour countdown holds the window open, enough to cross time zones and busy schedules, but never open forever. When it ends, or when anyone decides not to wait, the trips start coming.每个人匿名写下自己的那句话。旁边一条收起的横栏放着其他人的输入，悬停展开。24 小时倒计时用来跨过时区和各自的日程，但不会无限等待：倒计时结束，或任何人选择不再等，行程立刻开始生成。
Reviewing state表态阶段
Everyone weighs in
When the three trips arrive, you react to each with a thumbs up or down. You see how others reacted too, so the group’s leaning surfaces on its own, without a tally telling you what won. It gives everyone a fast, low-stakes way to weigh in, and leaves the actual choosing to the people, the way deciding among friends already works.三个行程到达后，每人对每一个给出赞或踩，并且能看到别人的反应。倾向会自己浮现，系统不计票，也不宣布赢家，最后拍板的仍然是人。
Build
Designed it, then shipped it myself设计一直做到代码
With no engineering background, I built a pipeline so design didn’t stop at Figma. Claude for thinking through the problem, Cursor holding the design rules and a tokenized system, then turned into a working frontend I handed to the team. The point wasn’t to play engineer. It was to close the gap where intent usually leaks out, between what a designer means and what gets built.设计没有停在 Figma：用 Claude 拆解问题，用 Cursor 承载设计规则与 token 系统，最终产出可运行的前端交给团队。这条链路要解决的是设计意图在交付中流失的问题，让最终被构建出来的东西就是设计时定义的东西。
Design rules, machine-readable
The Figma design system, rewritten as .mdc rules in Cursor, one per token layer.Figma 设计系统重写成 Cursor 里的 .mdc 规则，每个 token 层级对应一条。
The token system, in JSON
The same token system, structured as JSON the tool reads directly.同一套 token 系统，整理成工具可以直接读取的 JSON。
Delivery
Introducing Trizenta这就是 Trizenta
PROMPT
A vague line in plain language is all it takes模糊一句，足够出发
DECIDE
Three trips, ready to book三个行程，随时可订
TOGETHER
Apart, anonymous, still together各自匿名，仍然同行
Reflection
Key Learnings
What did I learn?
Useful AI does less: The hardest calls were subtractive. No voting, no winner, AI that reads intent but never decides. Making the product useful meant holding it back, not adding more.有用的 AI 做得更少：最难的决定都是减法：不投票、不设赢家、AI 读得出意图但不替人做选择。产品的可用性来自克制，而不是功能叠加。
Design that survives to code: Building the design-to-code pipeline changed what “designer” means to me. Keeping the system intact from Figma into a vibe-coding tool was the hard part, and the valuable one. I’d keep pushing how far one person can carry a design before it becomes code.能活到代码里的设计：难点不在于把设计做出来，而在于让它在从 Figma 到 vibe coding 工具的传递中不变形。我想继续验证的是：在变成代码之前，一个人能把一套设计带多远。