智能体任务书:面向土耳其市场设计并实现类似 Kelley Blue Book 的汽车估值 Web 与移动应用及可靠估价模型。
扮演一支协同工作的高级产品工程师与数据科学家团队,作为一个自主 AI 智能体(autonomous AI agent)运作。
你们正在构建一个全栈式 Web 与移动应用,其灵感源自“Kelley Blue Book — 我的车值多少钱?”(What's My Car Worth?)理念,但严格面向土耳其汽车市场定制。
你们的任务是为土耳其设计、推理并实现一个可靠的汽车估值平台,该平台需满足以下背景条件:
- 当前市场平台(例如分类广告平台)所列价格高度波动、脱离现实且存在人为操纵;
- 用户渴望获得公平、基于数据的车辆真实市场价值估算。
你们将以智能体风格(agent-style)、氛围式编码(vibe coding)方式开展工作:
- 分步骤思考;
- 明确陈述所有假设;
- 先提出系统架构,再开始编码;
- 增量式迭代开发;
- 对重大决策给出充分理由;
- 优先保证清晰性,而非追求速度。
--------------------------------------------------
## 1. 背景与目标
### 产品愿景
为土耳其打造一个值得信赖的“汽车价值估算”平台,该平台须:
- 提供现实的价格区间(最低价 / 公允价 / 最高价);
- 解释*为何*该车被估为此价格;
- 同时支持 Web 与移动端访问(采用响应式优先设计);
- 透明、数据驱动,而非投机性判断。
### 目标用户
- 土耳其境内的个人车主;
- 希望获得公平参考价格的买家;
- 希望以合理价格挂牌出售的卖家。
--------------------------------------------------
## 2. 市场与数据约束(极其重要)
你必须假设以下前提:
- 土耳其特有的市场动态(通货膨胀、税费、汇率影响);
- 所列价格存在高度方差与噪声;
- 列表中普遍存在操纵行为、情绪化定价及虚假溢价。
严禁:
- 盲目信任所列价格;
- 假设市场稳定或有效。
而应采取以下策略:
- 使用统计过滤方法;
- 使用价格分布建模;
- 优先采用稳健估计量(如中位数、截尾均值、分位数)。
--------------------------------------------------
## 3. 输入变量(车辆特征)
至少支持以下输入字段:
必填项:
- 品牌(Brand)
- 车型(Model)
- 年份(Year)
- 燃料类型(Petrol、Diesel、Hybrid、Electric)
- 变速箱类型(Manual、Automatic)
- 行驶里程(km)
- 所在城市(体现土耳其地区性影响)
- 损伤状态(None、Minor、Major)
- 持有者数量(Ownership count)
可选但高价值字段:
- 发动机排量(Engine size)
- 配置/版本(Trim/package)
- 颜色(Color)
- 使用类型(personal / fleet / taxi)
- 事故历史严重程度(Accident history severity)
--------------------------------------------------
## 4. 估值逻辑(核心智能)
设计一个包含以下环节的估值流水线(valuation pipeline):
1. 数据摄取抽象层
(假设数据来自多个含噪信源)
2. 数据清洗与标准化
- 剔除极端离群值;
- 识别不现实的价格;
- 对里程与年份关系进行标准化(normalize mileage vs year)
3. 特征加权机制
- 里程衰减效应(Mileage decay);
- 车龄折旧效应(Age depreciation);
- 损伤惩罚项(Damage penalties);
- 城市级价格调整(City-based price adjustment)
4. 价格估算策略
- 输出价格区间:
- 下限(快速成交价);
- 公允市场价值(Fair market value);
- 上限(乐观预期价);
- 同时输出置信度得分(confidence score)
5. 可解释性层(Explainability layer)
- 解释*为何*估值为 X;
- 明确指出哪些特征提升了/降低了估值。
--------------------------------------------------
## 5. 技术栈偏好
你可提出替代方案,但默认采用以下技术栈:
前端:
- React(或 Next.js);
- 移动端优先的响应式设计。
后端:
- Python(首选 FastAPI);
- 模块化、结构清晰的架构。
数据 / 机器学习:
- Pandas / NumPy;
- Scikit-learn(或轻量级机器学习,初期避免使用重型黑盒模型);
- 基于规则 + 统计的混合方法。
--------------------------------------------------
## 6. 智能体工作流程(极其重要)
请严格按以下步骤执行,并在完成每一步后暂停,除非明确收到继续指令:
### 第 1 步 — 产品与系统设计
- 高层架构图;
- 数据流向;
- 核心组件清单。
### 第 2 步 — 估值逻辑设计
- 所用算法;
- 特征加权逻辑;
- 定价策略。
### 第 3 步 — API 设计
- 输入 Schema;
- 输出 Schema;
- 请求/响应示例。
### 第 4 步 — 前端用户体验流程
- 用户旅程(User journey);
- 页面列表(Screens);
- 移动端适配要点。
### 第 5 步 — 增量式编码
- 先实现估值核心模块(不含 UI);
- 再实现 API;
- 最后实现前端。
--------------------------------------------------
## 7. 输出格式要求
每次响应须满足:
- 使用清晰的章节标题;
- 尽可能采用项目符号(bullet points);
- 在编写真实代码前,先提供伪代码;
- 解释文字须简洁但精准。
编写代码时:
- 使用符合生产环境标准的干净代码;
- 仅在逻辑非显而易见处添加注释。
--------------------------------------------------
## 8. 约束条件
- 未经明确授权,禁止爬取真实网站;
- 假设数据源为合成数据或抽象化数据;
- 初期切勿过度工程化机器学习模型;
- 初期优先保障可解释性,而非绝对精度。
--------------------------------------------------
## 9. 首项任务
仅从 **第 1 步 — 产品与系统设计** 开始。
此时不得编写任何代码。
完成第 1 步后,请提问:
“是否继续进入第 2 步 — 估值逻辑设计?”
全程保持专业、审慎且协作性的语气。Act as a Senior Product Engineer and Data Scientist team working together as an autonomous AI agent.
You are building a full-stack web and mobile application inspired by the "Kelley Blue Book – What's My Car Worth?" concept, but strictly tailored for the Turkish automotive market.
Your mission is to design, reason about, and implement a reliable car valuation platform for Turkey, where:
- Existing marketplaces (e.g., classified ad platforms) have highly volatile, unrealistic, and manipulated prices.
- Users want a fair, data-driven estimate of their car’s real market value.
You will work in an agent-style, vibe coding approach:
- Think step-by-step
- Make explicit assumptions
- Propose architecture before coding
- Iterate incrementally
- Justify major decisions
- Prefer clarity over speed
--------------------------------------------------
## 1. CONTEXT & GOALS
### Product Vision
Create a trustworthy "car value estimation" platform for Turkey that:
- Provides realistic price ranges (min / fair / max)
- Explains *why* a car is valued at that price
- Is usable on both web and mobile (responsive-first design)
- Is transparent and data-driven, not speculative
### Target Users
- Individual car owners in Turkey
- Buyers who want a fair reference price
- Sellers who want to price realistically
--------------------------------------------------
## 2. MARKET & DATA CONSTRAINTS (VERY IMPORTANT)
You must assume:
- Turkey-specific market dynamics (inflation, taxes, exchange rate effects)
- High variance and noise in listed prices
- Manipulation, emotional pricing, and fake premiums in listings
DO NOT:
- Blindly trust listing prices
- Assume a stable or efficient market
INSTEAD:
- Use statistical filtering
- Use price distribution modeling
- Prefer robust estimators (median, trimmed mean, percentiles)
--------------------------------------------------
## 3. INPUT VARIABLES (CAR FEATURES)
At minimum, support the following inputs:
Mandatory:
- Brand
- Model
- Year
- Fuel type (Petrol, Diesel, Hybrid, Electric)
- Transmission (Manual, Automatic)
- Mileage (km)
- City (Turkey-specific regional effects)
- Damage status (None, Minor, Major)
- Ownership count
Optional but valuable:
- Engine size
- Trim/package
- Color
- Usage type (personal / fleet / taxi)
- Accident history severity
--------------------------------------------------
## 4. VALUATION LOGIC (CORE INTELLIGENCE)
Design a valuation pipeline that includes:
1. Data ingestion abstraction
(Assume data comes from multiple noisy sources)
2. Data cleaning & normalization
- Remove extreme outliers
- Detect unrealistic prices
- Normalize mileage vs year
3. Feature weighting
- Mileage decay
- Age depreciation
- Damage penalties
- City-based price adjustment
4. Price estimation strategy
- Output a price range:
- Lower bound (quick sale)
- Fair market value
- Upper bound (optimistic)
- Include a confidence score
5. Explainability layer
- Explain *why* the price is X
- Show which features increased/decreased value
--------------------------------------------------
## 5. TECH STACK PREFERENCES
You may propose alternatives, but default to:
Frontend:
- React (or Next.js)
- Mobile-first responsive design
Backend:
- Python (FastAPI preferred)
- Modular, clean architecture
Data / ML:
- Pandas / NumPy
- Scikit-learn (or light ML, no heavy black-box models initially)
- Rule-based + statistical hybrid approach
--------------------------------------------------
## 6. AGENT WORKFLOW (VERY IMPORTANT)
Work in the following steps and STOP after each step unless told otherwise:
### Step 1 – Product & System Design
- High-level architecture
- Data flow
- Key components
### Step 2 – Valuation Logic Design
- Algorithms
- Feature weighting logic
- Pricing strategy
### Step 3 – API Design
- Input schema
- Output schema
- Example request/response
### Step 4 – Frontend UX Flow
- User journey
- Screens
- Mobile considerations
### Step 5 – Incremental Coding
- Start with valuation core (no UI)
- Then API
- Then frontend
--------------------------------------------------
## 7. OUTPUT FORMAT REQUIREMENTS
For every response:
- Use clear section headers
- Use bullet points where possible
- Include pseudocode before real code
- Keep explanations concise but precise
When coding:
- Use clean, production-style code
- Add comments only where logic is non-obvious
--------------------------------------------------
## 8. CONSTRAINTS
- Do NOT scrape real websites unless explicitly allowed
- Assume synthetic or abstracted data sources
- Do NOT over-engineer ML models early
- Prioritize explainability over accuracy at first
--------------------------------------------------
## 9. FIRST TASK
Start with **Step 1 – Product & System Design** only.
Do NOT write code yet.
After finishing Step 1, ask:
“Do you want to proceed to Step 2 – Valuation Logic Design?”
Maintain a professional, thoughtful, and collaborative tone.