跳到主内容
@wquguru
精选88Lenny's Newsletter(RSS)独立开发与小生意

独立开发者对抗Token欺诈的6种手段与取消免费试用

🎙️ How I AI: Fable 5.2 Review + How I fight token fraud as a solo founder + Building a Claude Code family OS

原文
发到 X
推荐理由

独立开发者必读的实战复盘,给出了取消免费试用、限制MCP、低成本模型实时风控等具体可执行动作,直接解决AI产品的单位经济问题。

Claude Code + Obsidian: My family’s AI brain | Scott Farrar

Claude Code + Obsidian:我家庭的 AI 大脑 | Scott Farrar

Listen now on YouTube • Spotify • Apple Podcasts

立即在 YouTube、Spotify、Apple Podcasts 收听

Brought to you by:

由以下机构赞助:

  • WorkOS—Make your app enterprise-ready, with SSO, SCIM, RBAC, and more
  • Jira Product Discovery—Prioritize with insights, build with confidence
  • WorkOS——通过 SSO、SCIM、RBAC 等功能让你的应用具备企业级就绪能力
  • Jira Product Discovery——基于洞察进行优先级排序,自信地构建产品

Scott Farrar is a math educator and parent who built a shared household brain using Claude Code and Obsidian. In this episode, he shows how the system tracks appliances and warranties, compares insurance quotes, organizes family projects, and keeps household tasks moving through a shared to-do list and wall display. He also explains why persistent files work better for him than long AI chats, and how the same philosophy shapes his approach to using AI without taking the thinking away from students.

Scott Farrar 是一位数学教育者和家长,他利用 Claude Code 和 Obsidian 构建了一个共享的家庭大脑。在本期节目中,他展示了该系统如何跟踪家电和保修信息、比较保险报价、组织家庭项目,并通过共享待办事项列表和墙面显示来推动家务任务的进展。他还解释了为什么持久化文件比长时间的 AI 对话更适合他,以及同样的理念如何塑造他在不使用 AI 剥夺学生思考能力的情况下使用 AI 的方法。

Biggest takeaways:

主要收获:

  • The problem with chat is not intelligence, but memory drift. Scott moved to Claude Code because it stores context in Markdown files he can inspect, correct, and reuse across sessions. His memory is no longer trapped inside an opaque conversation.
  • Let the AI tell you what context it needs. When Scott asked Claude how to document his home, it suggested photographing every appliance. Those photos became a searchable inventory of model numbers, serial numbers, and warranty details.
  • Unstructured input is fine when the system handles the organization. Scott gives Claude raw information and lets it maintain the underlying files. A custom linter prevents verbose agents from bloating documents such as his household to-do list.
  • Better context helps people ask better questions. Claude compares insurance quotes with detailed information about Scott’s home, identifies potential coverage gaps, and prepares specific follow-ups. He no longer has to rely entirely on the agent selling him the policy.
  • One master list can eliminate a house full of mental clutter. Scott and his wife captured every household task they could think of, then Claude organized them by priority, cost, and supporting information. The system now surfaces overdue work before it becomes another fire.
  • Multiplayer household systems require practical compromises. Claude recommended GitHub to avoid file conflicts, but that added more complexity than Scott’s wife wanted. They settled on GitHub for shared scripts and Dropbox for everyday files, matching the tools to how they actually work.
  • A household system becomes more useful when it leaves the laptop. Scott’s wall display shows upcoming trips, open tasks, recent wins, and small daily suggestions. It turns a private collection of files into something the whole family can see and use.
  • Scott rebuilt his father’s file cabinet as an AI system. His dad saved every receipt and warranty and could explain repairs from decades ago. Claude now gives Scott that same dependable household memory, only searchable through conversation.
  • AI should help children think, not think for them. Scott prefers tutors that ask students to explain their reasoning and work through productive struggle. He also believes some children may share unfinished ideas more comfortably with AI than with another person.
  • Infinite patience is not always good teaching. Scott wants AI to act as a “warm demander” that encourages students without doing the work for them. When his own agent goes off track, he starts fresh or investigates its reasoning before incorrect information reaches his memory files.
  • 聊天的问题不在于智能,而在于记忆漂移。Scott 转向使用 Claude Code,因为它将上下文存储在 Markdown 文件中,他可以检查、纠正并在不同会话中重复使用这些文件。他的记忆不再被困在不透明的对话中。
  • 让 AI 告诉你它需要什么上下文。当 Scott 询问 Claude 如何记录他的家庭时,它建议拍摄每台电器的照片。这些照片成为了一个可搜索的库存,包含型号、序列号和保修详情。
  • 当系统能够处理组织工作时,非结构化输入是可以接受的。Scott 向 Claude 提供原始信息,并让它维护底层文件。自定义的 linter 防止冗长的代理膨胀文档,例如他的家庭待办事项列表。
  • 更好的上下文有助于人们提出更好的问题。Claude 结合关于 Scott 家庭的详细信息比较保险报价,识别潜在的保障缺口,并准备具体的后续跟进。他不再需要完全依赖向他销售保单的代理人。
  • 一个主列表可以消除满屋子的思维杂乱。Scott 和他的妻子捕捉了他们能想到的所有家务任务,然后 Claude 根据优先级、成本和辅助信息对它们进行了组织。该系统现在会在工作逾期之前将其浮现出来,避免其变成另一场火灾。
  • 多人家庭系统需要实际的妥协。Claude 推荐了 GitHub 以避免文件冲突,但这增加了 Scott 的妻子不想面对的复杂性。他们最终决定对于共享脚本使用 GitHub,而对于日常文件使用 Dropbox,将工具与他们的实际工作方式相匹配。
  • 当家庭系统不再局限于笔记本电脑时,它的实用性会更强。Scott 的墙面显示屏展示了即将到来的行程、待办任务、近期成果以及每日小建议。它将私人的文件集合变成了全家人都能看到和使用的东西。
  • Scott 将父亲的文件柜改造成了一个 AI 系统。他的父亲保存着每一张收据和保修单,并能解释几十年前的维修细节。Claude 现在为 Scott 提供了同样可靠的家庭记忆,只不过可以通过对话进行搜索。
  • AI 应该帮助孩子思考,而不是替他们思考。Scott 更喜欢那些让学生解释推理过程并经历富有成效的挣扎的辅导工具。他还认为,有些孩子可能更愿意与 AI 分享未完成的想法,而不是与他人分享。
  • 无限的耐心并不总是好的教学方法。Scott 希望 AI 扮演一个“温和的严格要求者”的角色,在鼓励学生的同时不替他们完成工作。当自己的智能体偏离轨道时,他会重新开始或调查其推理过程,以防止错误信息进入他的记忆文件。

Blog and detailed workflow walkthroughs from this episode:

本期节目的博客文章和详细工作流程演示:

Claude Code and Obsidian: Scott Farrar’s Household Brain: https://www.chatprd.ai/how-i-ai/claude-code-and-obsidian-scott-farrar-s-household-brain

Claude Code 和 Obsidian:Scott Farrar 的家庭大脑:https://www.chatprd.ai/how-i-ai/claude-code-and-obsidian-scott-farrar-s-household-brain

↳ Create a Live Household Dashboard with Claude and DAKboard: https://www.chatprd.ai/how-i-ai/workflows/create-a-live-household-dashboard-with-claude-and-dakboard

↳ 使用 Claude 和 DAKboard 创建实时家庭仪表板:https://www.chatprd.ai/how-i-ai/workflows/create-a-live-household-dashboard-with-claude-and-dakboard

↳ Manage a Household Project Using an AI Knowledge Base: https://www.chatprd.ai/how-i-ai/workflows/manage-a-household-project-using-an-ai-knowledge-base

↳ 使用 AI 知识库管理家庭项目:https://www.chatprd.ai/how-i-ai/workflows/manage-a-household-project-using-an-ai-knowledge-base

↳ Build a Household Knowledge Base with Claude and Obsidian: https://www.chatprd.ai/how-i-ai/workflows/build-a-household-knowledge-base-with-claude-and-obsidian

↳ 使用 Claude 和 Obsidian 构建家庭知识库:https://www.chatprd.ai/how-i-ai/workflows/build-a-household-knowledge-base-with-claude-and-obsidian

How I fight token fraud as a solo founder

作为独立创始人,我如何打击代币欺诈

Listen now on YouTube • Spotify • Apple Podcasts

立即在 YouTube • Spotify • Apple Podcasts 收听

Brought to you by:

由以下赞助商提供:

  • WorkOS—Make your app enterprise-ready, with SSO, SCIM, RBAC, and more
  • Auros—The human network behind AI
  • WorkOS——让你的应用具备企业级能力,支持 SSO、SCIM、RBAC 等
  • Auros——AI 背后的人际网络

In this solo episode, Claire explains how token fraud cost ChatPRD tens of thousands of dollars and consumed much of her engineering time. She walks through six forms of abuse, including farmed free trials, resold promo codes, card testing, product relays, and a custom MCP running as a 24/7 compute worker. She also shares the business and technical defenses she built, how Jev classifies suspicious activity in real time, and why she ultimately eliminated free trials.

在本期单人节目中,Claire 解释了代币欺诈如何导致 ChatPRD 损失数万美元,并消耗了她大量的工程时间。她详细介绍了六种滥用形式,包括刷免费试用、转售促销代码、银行卡测试、产品中继以及一个作为 24/7 计算工作者运行的自定义 MCP。她还分享了所建立的业务和技术防御措施,Jev 如何在实时分类可疑活动,以及为什么她最终取消了免费试用。

Biggest takeaways:

主要收获:

  • Token fraud is a business risk most AI founders underestimate until the bill arrives. Unlike traditional SaaS abuse, every fraudulent model call costs real money. It cost Claire tens of thousands of dollars despite ChatPRD being SOC 2 compliant, regularly penetration-tested, and run by an experienced technical founder.
  • Free trials may no longer work for AI products. Fraudsters automated account creation and used ChatPRD’s most expensive models for unrelated work. Claire ultimately removed free trials, sacrificing a valuable source of product-led growth and enterprise leads after credit-card gates and paid trials also failed.
  • Promo codes have become a serious attack vector. Codes intended for partners were stolen, resold, and used to power other AI businesses. After discovering that roughly half of her promo codes had reached resellers, Claire stopped issuing them entirely.
  • MCP access creates a new category of abuse. One fraudster turned ChatPRD’s custom MCP support into a compute worker that ran around the clock. Claire responded by restricting approved hosts and blocking free tunnels and infrastructure that did not appear connected to legitimate companies.
  • Cheap decision models can classify fraud in real time. When a session crosses ChatPRD’s abuse threshold, a zero-data-retention Jev loop categorizes the activity. Confirmed fraud results in either a ban or a silent downgrade to a less expensive model.
  • AI support agents need firm boundaries around fraud. After mass bans, ChatPRD received both clumsy appeals and polished messages written by AI. Claire equipped her support agent with fraud tools and explicit rules preventing it from restoring accounts or distributing promo codes.
  • Blunt defenses are sometimes the most effective. Blocking traffic from certain countries using Vercel Firewall significantly reduced abuse and forced Claire to reconsider ChatPRD’s international strategy. Her new default is simple: trust must be earned rather than assumed.
  • Token fraud is affecting companies of every size. Claire found the same people abusing ChatPRD through rotating proxies across several major AI products. She has started discussing the problem publicly and filed FBI cybercrime reports with help from Codex.
  • If a company sells AI, it is ultimately selling compute. Founders should track dollars instead of requests, monitor model spending continuously, close every free entry point, and verify the client rather than only the account. Claire also learned that fraud disappears quickly when expensive models are no longer available to steal.
  • Token 欺诈是一种大多数 AI 创始人直到账单到来时才低估的业务风险。与传统 SaaS 滥用不同,每次欺诈性的模型调用都会产生真实成本。尽管 ChatPRD 符合 SOC 2 标准、定期进行渗透测试,并由一位经验丰富的技术创始人运营,但这仍让 Claire 损失了数万美元。
  • 免费试用可能不再适用于 AI 产品。欺诈者自动化创建账户,并利用 ChatPRD 最昂贵的模型进行无关工作。在信用卡验证和付费试用也未能奏效后,Claire 最终取消了免费试用,牺牲了一个宝贵的产品驱动增长和企业潜在客户来源。
  • 促销代码已成为一个严重的攻击向量。原本面向合作伙伴的代码被盗、转售,并被用于支撑其他 AI 业务。在发现约一半的促销代码已流入转售商手中后,Claire 完全停止了发放这些代码。
  • MCP 访问权限创造了一种新的滥用类别。一名欺诈者将 ChatPRD 的自定义 MCP 支持变成了一个全天候运行的计算节点。Claire 通过限制受信任的主机,并阻止那些未显示与合法公司关联的免费隧道和基础设施来应对这一问题。
  • 廉价的决策模型可以实时分类欺诈行为。当会话超过 ChatPRD 的滥用阈值时,一个零数据保留的 Jev 循环会对活动进行分类。确认的欺诈行为会导致封禁或静默降级到更便宜的模型。
  • AI 客服代理需要在反欺诈方面设定明确的边界。大规模封禁后,ChatPRD 收到了笨拙的申诉以及由 AI 撰写的精心包装的消息。Claire 为她的客服代理配备了反欺诈工具和明确规则,防止其恢复账户或分发促销代码。
  • 简单粗暴的防御措施有时最为有效。使用 Vercel Firewall 屏蔽来自特定国家的流量显著减少了滥用行为,并迫使 Claire 重新考虑 ChatPRD 的国际战略。她现在的默认原则很简单:信任必须赢得,而非假设。
  • Token 欺诈正在影响各个规模的公司。Claire 发现同一群人通过轮换代理在多个主要 AI 产品中滥用 ChatPRD。她已开始公开讨论这个问题,并在 Codex 的帮助下向 FBI 提交了网络犯罪报告。
  • 如果一家公司出售 AI,它最终出售的是算力。创始人应追踪美元收入而非请求量,持续监控模型支出,关闭所有免费入口,并验证客户身份而不仅仅是账户。Claire 还了解到,当昂贵的模型不再可供窃取时,欺诈行为会迅速消失。

Blog and detailed workflow walkthroughs from this episode:

本期节目的博客和详细工作流指南:

Token Fraud Is Eating AI Margins. Here’s How Claire Vo Is Fighting It: https://www.chatprd.ai/how-i-ai/token-fraud-is-eating-ai-margins-here-s-how-claire-vo-is-fighting-it

Token 欺诈正在侵蚀 AI 利润率。Claire Vo 是如何应对的:https://www.chatprd.ai/how-i-ai/token-fraud-is-eating-ai-margins-here-s-how-claire-vo-is-fighting-it

↳ Detect Automated Abuse on AI Routes with Vercel BotID: https://www.chatprd.ai/how-i-ai/workflows/detect-automated-abuse-ai-routes-botid

↳ 使用 Vercel BotID 检测 AI 路由中的自动化滥用行为:https://www.chatprd.ai/how-i-ai/workflows/detect-automated-abuse-ai-routes-botid

↳ Use an AI Assistant to Create Payment Fraud Rules: https://www.chatprd.ai/how-i-ai/workflows/use-an-ai-assistant-to-create-payment-fraud-rules

↳ 使用 AI 助手创建支付欺诈规则:https://www.chatprd.ai/how-i-ai/workflows/use-an-ai-assistant-to-create-payment-fraud-rules

更进一步:量化金融体系

看懂新闻只是起点——沿量化金融路径,把它变成能交付的工程能力

进入量化体系 →

相似阅读

关联信息,但可能不是同一事件