动手教程 · 给认真的买家和卖家A HANDS-ON GUIDE · FOR SERIOUS BUYERS & SELLERS
动手做一个你自己的 AI 地产助手——也看清它的能力边界 Build your own AI real estate agent — and see exactly where it falls short
用现成的免费工具，大约 20 分钟就能搭一个属于你的看房助手。然后我会带你看清：它能帮你到哪一步，又在哪里仍然需要一个有经验的真人。 With free, no-code tools, you can build a property assistant of your own in about 20 minutes. Then I’ll show you how far it gets you — and where it still needs an experienced human.
什么是 AI real estate agent？What is an AI real estate agent?
AI real estate agent 是一个用大语言模型（如 Claude、GPT、Gemini）搭起来、专门回答房产问题的对话助手。你可以让它整理信息、比较选项、解释买卖租的流程。但它不持牌、不能代表你签约，并且在缺乏实时数据时会”编”出听起来合理、实则错误的答案——这在 AI 研究里有个公认的名字，叫”幻觉”（hallucination）。 An AI real estate agent is a conversational assistant built on a large language model (like Claude, GPT, or Gemini) and pointed at real-estate questions. You can have it organize information, compare options, and explain the buying, selling, or renting process. But it is not licensed and cannot represent you in a contract, and when it lacks live data it will “invent” answers that sound reasonable but are wrong — a well-documented behaviour that AI research calls hallucination.
第一部分 · 动手PART ① · BUILD IT
20 分钟：做一个你自己的20 minutes: build one yourself
不用写一行代码。四步就有一个能回答房产问题的助手。No code, four steps to a working real-estate Q&A assistant.
选一个平台Pick a platform
挑一个你能直接上手的：Claude Projects、ChatGPT 的自定义 GPT、Gemini Gem，或 NotebookLM。它们都能”读”你给的资料、按你的要求回答。免费版起步就够。 Pick one you can use today: Claude Projects, a ChatGPT Custom GPT, a Gemini Gem, or NotebookLM. Each can “read” the material you give it and answer the way you ask. The free tiers are enough to start.
四个平台怎么选，看我的对比 →How to choose between them →喂它”知识”Feed it knowledge
把你关心的资料丢进去：你在看的社区、收藏的房源、关心的预算和学区、你反复想问的问题。这一步决定它有多”懂行”——喂得越薄，它在没资料时越容易编。 Give it the material you care about: the neighbourhoods you’re watching, listings you’ve saved, your budget and school priorities, the questions you keep asking. This step decides how informed it is — the thinner you feed it, the more it invents when it runs out of facts.
写”人设”指令Write its instructions
用一段话告诉它怎么当你的助手：当一个 GTA 买家助手、先问我的预算和时间线、绝不编造价格、不确定就说不确定、回答尽量标出处。这几句”卫生习惯”能挡掉一大半离谱回答。 In a short paragraph, tell it how to behave: act as a GTA buyer’s assistant, ask about my budget and timeline first, never invent prices, say “I’m not sure” when it isn’t, and cite sources where it can. These few hygiene rules block a large share of nonsense answers.
拿我整理的 Prompt 库做起点 →Start from my prompt library →测试它，找裂缝Test it — find the cracks
拿真实问题考它：某条街最近成交多少、这套房合不合理、学区边界在哪、昨天新挂的那套怎么样。你会很快发现它的边界——这正是下一部分要讲的：它在哪里靠不住，为什么。 Quiz it with real questions: what sold on that street, is this listing fairly priced, where’s the school boundary, what about the one listed yesterday. You’ll find its limits fast — which is exactly the next part: where it can’t be trusted, and why.
第二部分 · 现实检验PART ② · REALITY CHECK
你的 AI 助手会在哪里出错Where your AI assistant breaks
建出来之后你会发现：它整理信息、解释流程很好用，但一碰到”此时此地的真实数据”就开始失准。两个最常见的裂缝：Once you’ve built one, you’ll notice: it’s great at organizing information and explaining process, but it loses accuracy the moment it touches “real data, here and now.” The two most common cracks:
一个只靠通用模型的 DIY 助手，可能给你一个听起来很具体、其实是猜的数字——它没有实时 MLS 数据，于是”补全”成一个看似合理的价位。这就是幻觉：自信，但可能错。A DIY assistant on a general model may hand you a number that sounds precise but is guessed — it has no live MLS data, so it “fills in” a plausible figure. That’s hallucination: confident, possibly wrong.
一个接地的专业 agent 会说”我用的是截至某日的真实成交记录”，给出区间并附上每一笔的出处；拿不到数据时直接说拿不到，而不是猜。A grounded, professional agent says “here’s what actually sold, as of this date,” gives a range with a source for each comp, and tells you when it simply doesn’t have the data — instead of guessing.
上图为说明性示例，非真实成交记录。真实 comps 以持牌经纪调取的 MLS 数据为准。Illustrative example, not real sold records. Actual comps come from MLS data pulled by a licensed agent.
第二个裂缝：能力本身就有差别The second crack: capability itself varies
“AI 助手”不是一个东西。免费 Custom GPT，和一个接了真实数据、持续更新、回答带出处、背后有持牌人负责的 agent，是两个量级。“AI assistant” isn’t one thing. A free Custom GPT and an agent connected to real data, continuously updated, citing sources, with a licensed human behind it — are two different leagues.
| 维度Dimension | DIY 免费助手DIY free assistant | 专业 agent + 真人Pro agent + human |
|---|---|---|
| 数据准确性Data accuracy | 凭训练记忆，可能过时或编造From training memory; may be stale or invented | 接真实成交/挂牌数据，标出处Real sold/listing data, with sources |
| 实时 MLSLive MLS | 拿不到No access | 持牌经纪可调取Pulled by a licensed agent |
| 本地经验Local experience | 没有走过的房、谈过的 dealNo homes walked, no deals closed | 12 年全职、实地走过的判断12 years full-time, on-the-ground judgment |
| 法律合规Legal & compliance | 不持牌，不能给可依赖的合规意见Unlicensed; no reliable compliance advice | 持牌、受 RECO 监管下负责Licensed and accountable under RECO |
| 谈判Negotiation | 只能给通用话术Generic scripts only | 真实多方谈判经验Real multi-party negotiation |
| 问责Accountability | 出错没人负责No one is accountable for mistakes | 有受托责任，对结果负责Fiduciary duty; responsible for the outcome |
③ 想深入？专业级 agent 长什么样③ Going deeper: what a pro-grade agent looks like
我自己就给这个网站做了一个接地的 AI 助手。它不是裸跑的通用模型，而是：I built a grounded AI assistant for this very site. It isn’t a bare general model — it is:
而”专业级”不是写个 prompt——它背后是一整套工程：And “professional-grade” isn’t one prompt — it’s a whole stack of engineering:
- Prompt Engineering · 提示工程 写清楚它该怎么回答、什么不能做defining how it should answer, and what it must not do
- Context Engineering · 上下文工程 每次对话喂给模型哪些信息——决定成败的关键what the model sees each turn — where it’s won or lost
- Retrieval / RAG · 检索工程 接真实数据，让答案有出处、不靠猜grounding answers in real data, with sources — not guesses
- Data Engineering · 数据工程 把知识库整理干净、持续更新building and continuously updating a clean knowledge base
- Evaluation (Evals) · 评估工程 系统地测它答得对不对、有没有幻觉systematically testing whether it’s right — and catching hallucinations
- Guardrails · 安全护栏 挡住越界、合规与法律风险blocking out-of-bounds, compliance, and legal risks
- Orchestration · 编排 · 工具调用 让它能多步做事、调用工具letting it take multi-step actions and call tools
- Observability & Ops · 可观测 · 运维 上线后监控、追踪、迭代monitoring, tracing, and iterating once it’s live
想看它怎么造、踩过哪些坑，去我的 AI 学习手记：To see how it’s built and what I learned, visit my AI learning journal: Agent 进阶 → 日常项目 →Projects →
就在本站问问我的 AI →Ask my AI right here →第三部分 · 主张PART ③ · THE POINT
不是 AI vs 经纪，是 AI + 经纪It’s not AI vs. an agent — it’s AI + an agent
把上面的局限连起来看，结论不是”AI 没用”，而是分工变了。AI 擅长整理信息、跑初步分析、把流程讲清楚；而定价里的取舍、合同条款的风险、谈判桌上的临场判断，以及对最终结果负责——这些仍然来自一个走过很多套房、谈过很多笔 deal 的真人。Put those limits together and the takeaway isn’t “AI is useless” — it’s that the division of labour changed. AI is good at organizing information, running first-pass analysis, and explaining the process. The judgment in pricing, the risks buried in contract clauses, the read on a negotiation, and accountability for the outcome still come from a human who has walked many homes and closed many deals.
我把”哪些可以交给 agent、哪些永远需要人（法律、定价、合规）”整理成了一份清单——I’ve turned “what you can hand to an agent, and what always needs a human — legal, pricing, compliance” into a checklist — 可交给 Agent 的地产技能 →real-estate skills to hand to an agent →
接下来WHAT’S NEXT
你的 agent，和我的 agentYour agent, and mine
我在做的下一步：让你的个人 AI 助手和我的接地 agent 对接。你的 agent 管你的偏好和初筛，我的 agent 给出有真实数据和出处的答案，两者在中间对接，把繁琐的细节处理掉，把需要判断的留给人。What I’m building next: letting your personal AI assistant connect with my grounded agent. Yours handles your preferences and first-pass screening; mine supplies answers backed by real data and sources; the two meet in the middle to handle the tedious details and leave the judgment calls to people.
我也相信这是好事：当”透明、可核查、有出处”成为标准，靠信息差和话术生存的做法会越来越难——这对认真做事的人是顺风。I also think this is healthy: when transparent, verifiable, sourced answers become the norm, practices that rely on information gaps and sales talk get harder to sustain — a tailwind for anyone doing the work honestly.
常见问题FAQ
关于自建 AI 地产助手On building your own AI real estate agent
我真的能免费自己做一个 AI 地产助手吗？Can I really build my own AI real estate agent for free?
可以。用 Claude Projects、ChatGPT 自定义 GPT、Gemini Gem 或 NotebookLM 的免费档，约 20 分钟就能搭一个会回答房产问题的助手。它适合整理信息和学习流程，但不持牌、不能代表你交易。Yes. Using the free tier of Claude Projects, a ChatGPT Custom GPT, a Gemini Gem, or NotebookLM, you can build a working real-estate Q&A assistant in about 20 minutes. It’s good for organizing information and learning the process — but it isn’t licensed and can’t represent you in a transaction.
AI 助手会取代地产经纪吗？Will an AI agent replace my realtor?
不会取代，但会改变分工。AI 擅长整理和初步分析；定价、合规、谈判、对结果负责，仍需要持牌、有经验的真人。更现实的图景是”AI + 经纪”协作，而不是二选一。It won’t replace one, but it will change the division of labour. AI is good at organizing and first-pass analysis; pricing, compliance, negotiation, and being accountable for the outcome still need a licensed, experienced human. The realistic picture is “AI + agent,” not either-or.
为什么我的 ChatGPT 地产助手会报错价格？Why does my ChatGPT real estate assistant give wrong prices?
因为通用模型没有实时 MLS 数据。当它没有真实成交记录时，会”补全”出一个看似合理的数字——这就是幻觉。解决办法是给它接真实数据源，或用持牌经纪调取的 comps 核对。Because a general model has no live MLS data. When it lacks real sold records, it “fills in” a plausible-looking number — that’s a hallucination. The fix is to connect it to real data sources, or verify against comps pulled by a licensed agent.
免费 Custom GPT 和专业 AI agent 差在哪？What’s the difference between a free Custom GPT and a professional AI agent?
差在数据接地、实时性、出处可追溯、以及背后有没有持牌人负责。专业 agent 接真实数据、回答带来源、持续更新，并由持牌经纪对结论负责；免费 Custom GPT 三样都缺。It comes down to grounding in real data, freshness, traceable sources, and whether a licensed human stands behind it. A professional agent is connected to real data, cites sources, updates continuously, and a licensed agent is accountable for the conclusions; a free Custom GPT has none of those.
我的 AI 助手将来能和 Arthur 的 agent 对接吗？Could my AI assistant work with Arthur’s agent one day?
这正是我在做的方向：你的个人 AI 助手负责你的偏好和初筛，我的接地 agent 提供有出处的答案，两者对接处理细节。现在还在早期——想第一时间知道进展，可以在下方留个联系方式。That’s exactly the direction I’m building toward: your personal AI handles your preferences and first-pass screening, my grounded agent supplies sourced answers, and the two connect to handle the details. It’s early days — leave your contact below if you’d like to hear when it’s ready.
