小绿茶Xiaolvcha

OPC 创业OPC venture

我把一个会算命的聊天机器人做成了实体手办,我和我的小伙伴两个人经历了产品定义、系统开发、硬件设计、手办落地、电商运营、自媒体、投流、众筹……一起投钱深入把产品做闭环。

My partner and I turned a fortune-telling chatbot into a physical figure, moving through product definition, system development, hardware design, production, e-commerce, social media, paid acquisition, and crowdfunding. We invested our own money to close the full product loop.

实体成品实拍 Physical prototype, photographed in hand

01 Trigger

算命是极强的
启动场景

Fortune-telling creates an unusually strong trigger.

算命是最情绪化的场景,用户来到这里时,已经带着情绪、上下文和继续追问的动力。

Fortune-telling is intensely emotional. Users arrive with feelings, context, and a reason to keep asking questions.

产品痛点页 Product pain-point page

产品起点 Starting point

焦虑,是我们抓的第一痛点,问出你心中的最深的疑惑。

Anxiety was the first pain point we targeted: give people a way to ask their deepest question.

02 Accuracy

“算得准”的工程化手段
三层设计

A three-layer system engineered for accurate results.

通用模型的缺陷在计算能力,这是我们找到最讨巧的切入口:“大模型套上排盘工具,准得吓人”。大模型天然适合解盘,但不适合排盘。因此先接入八字、梅花易数等工具,再把结构化结果交给模型解释,最后用长期上下文维持角色判断。

General models were weak at deterministic calculation, which gave us a useful opening: pair the model with charting tools and the result becomes strikingly accurate. The model was naturally good at interpretation, while tools handled chart construction. Structured results then entered the model, with long-term context maintaining the character’s judgment.

  1. 01

    算对 Calculate

    工具完成排盘、起卦与确定性计算。

    Tools handle charting, divination, and deterministic computation.

  2. 02

    读懂 Interpret

    模型把结果放进用户关系、事件与历史对话里。

    The model connects results with the user’s relationships, events, and prior dialogue.

  3. 03

    站住 Hold

    角色在质疑面前维持判断,并在新证据出现时重新验证。

    The character holds a position under challenge and revalidates it when new evidence appears.

出生与问题Input 排盘工具Tool 结构化结果Data 模型解释LLM 长期上下文Memory

智能体后端系统 Agent operations backend

独立后台 Dedicated backend
这是小绿茶的独立后台,从小智项目开源修改,做了大量适配工作。This dedicated Xiaolvcha backend was adapted extensively from the open-source Xiaozhi project.

03 Boundary

算命场景不能
立刻滑跪,也不能
持续顶嘴

A fortune-telling agent cannot fold immediately or keep arguing.

立即认错会失去权威,持续嘴硬会放大错误。这是工程设计中最难的,真正的设计空间位于两者之间:先解释判断依据,再识别质疑强度,必要时重新调用工具,最后明确修正。

Immediate surrender weakened authority, while endless insistence amplified errors. This became the hardest engineering problem: explain the basis, read the strength of the challenge, re-run tools when needed, then correct the conclusion explicitly.

立刻认错Instant surrender
解释 核验 修正Explain verify correct
持续嘴硬Endless insistence

Vibe coding 实验 Adversarial build loop

两个 API 对打,
一轮一改 Prompt。

一个模型扮演小绿茶,另一个模型根据上一轮回答生成新的攻击。每轮记录角色是否站得住、工具是否调用、承诺是否执行、上下文是否被污染。

One model played Xiaolvcha; another generated the next challenge from the previous response. Each round tracked authority, tool use, promise execution, and context contamination.

  1. 01用户质疑Challenge
  2. 02小绿茶回应Response
  3. 03第二个模型追击Next attack
  4. 04分析失败模式Diagnose
  5. 05修改 Prompt 再跑Revise & rerun
测试材料目录:联调、压力、真实工具、性能与专业坚持机制。Test archive: integration, stress, real-tool, performance, and authority-calibration reports.
五个质疑场景:排盘数据、运势、性格、虚假工具声明与极端攻击。Five challenges: chart data, forecast, personality, false tool claims, and an extreme attack.
发现 01 Finding

“我重新算一下”必须在同一轮真的调用工具。

A promise to recalculate had to trigger the tool in the same turn.

发现 02 Finding

连续质疑会进入记忆,后面的“自信”可能来自被污染的上下文。

Repeated attacks entered memory, so later confidence could come from contaminated context.

发现 03 Finding

工具调用的惯性很短,下一轮仍需要清晰的触发条件。

Tool-use momentum was short-lived; each new turn still needed an explicit trigger.

04 Object

打造小绿茶的
身体

Building a body for Xiaolvcha.

服务器、ASR、TTS、LLM 与工具链先形成可说话的系统,随后接入设备端、PCB、外壳、三维拆件和样机。每一层都会反过来改变交互等待、角色表达与成本。

The server, ASR, TTS, LLM, and tool chain formed a speaking system before device logic, PCB, enclosure, 3D parting, and prototyping were added. Each layer changed latency, character expression, and cost.

三视图 打印 固化 白模 上色与装配

Turnaround print cure white model paint and assembly

内部组件实拍 PCB, battery, and wiring
功能结构与拆件 Functional structure and parting

从软件到设备 Software to device

硬件把每个抽象决定都变成实际的成本。

Hardware turned every abstract decision into a real cost.

产品介绍 Product story

从对话开始,剩下的情绪、焦虑、未来,都交给小绿茶。

Start with a conversation, then leave the emotion, anxiety, and future to Xiaolvcha.

产品定义 Proposition

算、卜、星、象,以及预测、陪伴、指引。

三个启动时刻 Three moments

深夜关系、出门仪式与工作决策。

能力延展 Extensions

更多日常问题、东西方术数切换、个人小程序、礼赠场景。

05 Design

产品越来越漂亮,
成本越来越重

The product kept getting more beautiful, and the cost kept getting heavier.

实体成品、白模、手持比例、三视图、结构、包装与商品表达形成了完整的设计链。精致的设计解决了表达问题,也把项目带进 IP 运营、供应链、库存与投流的更长链路。

Physical finishes, white models, hand-scale tests, turnarounds, structure, packaging, and commerce formed a complete design chain. Refined design solved the expression problem while extending the project into IP operations, supply chain, inventory, and paid distribution.

角色探索Character材质系统Material三维制造3D build商品表达Commerce

小绿茶的实体成品、手持比例、制造过程、结构设计、使用场景、包装与发布物料。点击任意图片可查看原图。

Xiaolvcha across physical finishes, hand-scale tests, making, structure, use contexts, packaging, and launch materials. Open any image for the full view.

06 Market

需求与产品形态的
偏离

The demand was real; the product form drifted away from it.

直到投流开始,我逐渐发现,痛点找对了只是第一环。我们抓的情绪化场景流量很大,私信多,但用户更多关注的是他的问题被快速回答。这种严肃算命的场景,用户等不及实体产品,我们的竞品偏离到真人师傅,从定价到产品形态都相对过重。

Paid acquisition gradually revealed that finding the pain point was only the first step. The emotional scenario attracted traffic and direct messages, while users cared most about getting an answer quickly. In serious fortune-telling, a physical product arrived too slowly; the real competitive reference shifted toward human practitioners, leaving both price and product form too heavy.

小红书曝光 Xiaohongshu30w+

自然流 + 聚光投流

Organic + Spotlight
最低互动成本 Lowest CPE≈ ¥0.15

小红书聚光

Xiaohongshu Spotlight
总体流量曝光 Total reach50w+

综合链路口径

Cross-channel total
众筹支持 Crowdfunding127 台

淘宝众筹

Taobao crowdfunding

01 热度 Attention

星座是流量收割机,赛道本身没问题,产品定义更需要看功力。

Astrology is a traffic magnet. The category has demand; product definition determines the outcome.

小绿茶 AI 国潮手办小红书主页截图 小绿茶 AI 国潮手办 · 泓垣Xiaohongshu profile · Opens in a new tab xiaohongshu.com

星座、塔罗与玄学内容能够获得互动,说明焦虑、关系与确定性这些问题本身具有传播张力。

Astrology, tarot, and divination content generated interaction, showing that anxiety, relationships, and certainty carried real distribution energy.

02 购买 Purchase

A3 人群到 A4 人群,决定了产品的生死线。

The move from A3 to A4 determined whether the product could survive.

我先围绕星座、考试、户外等场景建立小红书内容 SOP,再用聚光测试人群和关键词。交易侧接入淘宝众筹,并用万相台继续测试流量,最终获得 127 台众筹支持。把整体数据做成漏斗后,断层出现在 A3(询问)到 A4(行动):用户愿意问,购买动机还不够强。

I built a Xiaohongshu content SOP around astrology, exams, outdoor scenarios, and other topics, then used Spotlight to test audiences and keywords. Taobao crowdfunding and Wanxiangtai extended the experiment into purchase, resulting in support for 127 units. The funnel broke between A3 (Ask) and A4 (Act): users were willing to inquire, while purchase motivation remained weak.

复盘了众筹反馈、开模成本、投放成本与转化数据。严肃命理、全新角色和实体硬件叠在一起,让用户需要一次理解太多新东西。

The review covered crowdfunding feedback, tooling cost, media cost, and conversion data. Serious divination, a new character, and physical hardware asked users to understand too many new ideas at once.

三点反思 Three reflections

严肃命理需要工具与信任

它面对的是算命师傅和长期关系;塔罗拥有更轻的仪式、更直观的视觉和更短的体验链。

Serious divination relies on tools and trust, while tarot offers a lighter ritual, clearer visuals, and a shorter experience loop.

三种创新叠加了教育成本

AI 玩具、全新 IP 与新商品类目同时出现,用户理解、搜索、比较和下单都缺少现成参照。

An AI toy, a new IP, and a new retail category arrived together, leaving users without familiar references for discovery, comparison, or purchase.

投流衡量素材传播效率

内容热度说明问题有张力,交易仍需要清晰的拥有理由、价格锚点与持续运营。

Content heat showed that the question carried tension; a transaction still required a clear ownership reason, price anchor, and sustained operation.