何时打开:你要做创业 / 副业 / 评估别人项目 / 加入新公司,想用一个综合框架而不是单一作者视角。本 wiki 把 5 派方法论(yitang Truman 的五步法 / 王慧文清华产品课 / 王凯算账派 / 老黄独立产品 / 深思圈 VC 视角)抽出共识、显式分歧、各派独家维度。
一句话核心:没有"对"的创业方法论,只有"匹配你身份"的方法论。学生 / 大厂离职创业者 / 产品经理 / 工程师 / 独立开发者,各自走不同的路径。最大的反 pattern 是把方法论当 BP 工具用 —— 应该当"假设清单"用。
0. 5 派的创业观
| 派别 | 来源 | 核心问题 | 一句话立场 |
|---|---|---|---|
| 一堂 Truman 派 / 关键假设科学派 | 一堂机会预判方法论 / 一堂 / 五步法完整体系(关键假设 + 业务里程碑 + 五步法 5 进阶 + 落地) / 一堂 / 商业预判进阶(行业画布 + 集中度 + 案例) | "如何用一套科学方法把成功率从 10% 提升到 76%?" | "用假设驱动业务,五步法是关键假设的容器" |
| 王慧文清华产品课派 / 美团派 | 一堂 / 王慧文清华产品课 66 页完整笔记(Patrick 整理 / 2020) | "为什么失败是成功的 30 倍?" | "规模效应是商业世界万有引力 / 真正核心竞争力 = 发现机会 + 持续学习 / 运气占成功 90%" |
| 王凯算账派 / AI 时代小微 team | 王凯 AI 时代创业方法论与商业判断框架 | "怎么做生意赚现金流,不是融资?" | "互联网遗毒是融资 = 成功,我们要做生意 / 多项目并行 / 算账先于动手" |
| 老黄独立产品派 / 非技术全球商人 | newtype · Vibe Coding 独立产品实战 / newtype 心法 · 超级个体方法论 | "1 人 + 非技术 + 小钱起,怎么做独立产品?" | "Unit Economics + 把技术问题转成商业问题 + 强项做骨架" |
| 深思圈 VC 派 / a16z 红杉视角 | 深思圈作者全集入口 | "这家 AI 公司估值合不合理?" | "看 a16z / 红杉 / YC 怎么投 + 创始人靠不靠谱 + 公司案例库" |
关键观察:5 派的分歧不是对错之争,而是定位之争:
- yitang 教 MBA 学员 + 创业者(科学方法体系)
- 王慧文教清华学生(经典产品方法论 + 美团基因)
- 王凯写给独立开发者 / 小微 AI team(现金流生存)
- 老黄写给非技术个人(从 0 到独立产品)
- 深思圈写给关注 AI 行业的 PM / 投资人(VC 视角)
1. 跨作者共识(6 条)
1.1 "失败远多于成功",不是例外是默认
| 作者 | 数据/表述 |
|---|---|
| 王慧文 | "失败是成功的 30 倍" / 列了 5 对失败-成功(诺基亚 / iPhone — MPMan / iPod — Kik / 微信 — 智能 ABC / 搜狗 — IE / Chrome) |
| 王凯 | "上千个项目才出来一家成功上市,5 万人基数不到 100 人能落袋百万" / "项目大概率是要死的,90% 不咋赚钱" |
| yitang Truman | "创新型追求头部 < 3%,甚至 1‰" / 业务类型成功率分级(都赚钱副业 < 50% → 小餐馆 < 30% → 独立规模化 < 15% → 创新追求头部 < 3%) |
| 老黄 | "项目大概率失败 / 独立开发不要给 1 年耐心" |
| 王兴(王慧文转述) | "先胜而后求战(孙子兵法)/ 不要找一个自己打不赢的滩头阵地" |
共识结论:预判和准备的价值,在于把"小概率事件"变成"大概率能跑通"。
1.2 算账先于动手,这是科学创业的入场券
| 派 | 算账重点 | 关键数字/工具 |
|---|---|---|
| 王凯(三件事) | 流量获取成本 / 商业化效率 / 长期运维成本 | "单技术成本算账:全栈 3 万月薪 → 需要 13,334 个月付费用户 / 必须做稍微大点的项目" |
| yitang Truman(拆推评算 4 字诀) | 需求频率 + 规模 + 紧迫性 → 单元经济 → 天花板 | "没有定量,就没有真正的需求评估" |
| 老黄(Unit Economics 3 心法) | 掂量能耐和资源 / 把技术问题转成商业问题 / 从强项出发 | "不碰大型/复杂项目 / 客户不为代码漂亮买单" |
| 王慧文(美团基因) | 销售/用户量足够大之后(与小竞争对手比)产生的用户体验优势 + 成本优势 | "规模效应是商业世界万有引力" |
| 深思圈 | a16z / 红杉的"25× 收入倍数 + 5× 烧钱倍数" + Sequoia Agent Economy 三阶段 | VC 视角 |
共识结论:算账没算对是 AI 产品死法第一名(王凯亲历 6 个项目,5 个因为算账问题死了)。
1.3 预判 > 执行(选项池 / 终局 / 关键假设)
| 派 | 表述 |
|---|---|
| yitang Truman(一堂机会预判方法论) | "预判往往比一堂五步法重要数倍 / P 型 vs L 型成功率差 100 倍" |
| 王慧文 | "互联网不是内卷,是市场驱动 → 领导力驱动 → 创新驱动 3 阶段,现在是市场驱动结束阶段所以大家觉得难" / "真正的核心竞争力 = 发现机会的能力 + 持续学习的能力" |
| 王凯 | "新产品 90% 精力做推荐流量,SEO 是滞后指标" / 6 心法之一:"不挑战巨头主战场" |
| 老黄 | "从商业角度切入,能不能把技术问题转成商业问题?批发 API 再零售" |
共识结论:起盘前 4 周到 4 个月的预判,远比起盘后的执行重要。
1.4 5 步法 / 5 维度 框架(几乎所有派都用类似结构)
虽然命名不同,但 5 派对"从 0 到 1 要拆什么"高度一致:
| 维度 | yitang 五步法 | 王慧文产品课 | 王凯算账 | 老黄 |
|---|---|---|---|---|
| 需求 | ① 需求(拆推评算 4 字诀) | 需求 4 大特质 + 供需关系 | 找需要付费的人 | 客户买"需求点打得准" |
| 解决方案 | ② 解决方案(产品内核 + 单元模型) | 失败 vs 成功 30 倍 | 不挑战巨头主战场 | 把技术问题变商业 |
| 商业模式 | ③ 商业模式(单元经济) | 规模效应 ABC 类 + 网络效应 | 算账三件事 + ARPU vs 即时付费 | Unit Economics |
| 增长 | ④ 增长(LTV / 获客成本) | 创新的扩散 + 4P + STP + 高频打低频 | 流量红利演化 + SEO/SEM 攻防 | 强项做骨架(营销/运营/产品) |
| 壁垒 | ⑤ 壁垒(项目壁垒) | 波特 5 力 + 竞争 3 战略 + 美团平台战略 | AI 应用层三步走(用 AI 切入 → 攒规模 → 延伸非 AI 壁垒) | 强项 + 跨平台配置 + 业务决策 |
共识结论:这 5 个维度缺一不可,任何一个缺失都是"假动作"。
1.5 用"假设"替代"我认为"(Truman 灵魂拷问被多派承袭)
一堂 / 五步法完整体系(关键假设 + 业务里程碑 + 五步法 5 进阶 + 落地) §三:Ω 模型(提假设 → 验证假设 → 迭代/淘汰)。
5 问灵魂拷问(任何创业讨论都该过):
- 你的"我认为"背后,实际的客观假设是什么?
- 这个假设你打算怎么验证?
- 如果假设不成立,你的业务会怎样?
- 验证这个假设需要多少成本?
- 这个假设是 60% 把握 还是 80%?
沟通升级:"我认为..." → "我的假设是 X / 验证方式 Y / 现状 Z"。
王凯也认同:"AI 产品的死法第一名是鸵鸟心态——埋头做不思考流量、付费、扩张";老黄也认同:"先想清楚商业,Cursor 会帮你写"。
1.6 一定会死,所以要"多项目并行 / 多假设并行"
| 派 | 表述 |
|---|---|
| yitang(76% 概率数学) | "一年测 4 次 → 总成功率 P = 1 - 0.7⁴ = 76%" |
| 王凯(多项目并行) | "连续成功 5 个小项目的概率低于成功 1 个稍大项目" / "做完赚钱低成本维护,再做下一个,不要在 ROI 低的机会卷" |
| 王慧文 | "我们一开始只想做一个 10 亿美金市值、60 人的公司,没想到后来做大了" |
| 老黄 | newtype.pro + Prompt House + LearnFromYouTube + Super Analyst + newtype Profile + CLI + OS,6+ 产品并行 |
共识结论:单押一个项目 = 押小概率事件;并行多个项目 = 把小概率推到大概率。
2. 显式分歧(4 条 —— 不抹平)
2.1 分歧 1:要不要融资?
| 视角 | 立场 | 关键论据 |
|---|---|---|
| 王凯(反融资) | "做生意比融资重要" | "互联网时代极少创始团队真正赚到钱,反而生意老板年盈利几百万还能旅游打德州" / "融资成功只是上了牌桌,至于是不是废牌还没验证" / "几位 VC 朋友说 AI 创业首选北京,人才都在北京 → 我反驳:默认前提是大生意" |
| yitang Truman(看终局) | "保 A 争 B" | 7 层光谱图 / 选区间不是点 / 终局 1-2 是副业(零花钱)→ 终局 7 是上市(十亿级+),融资只是终局 4+ 才需要 |
| 王慧文(美团 No.2) | 融资是手段不是目的 | "我们一开始只想做 10 亿美金市值、60 人的公司,没想到后来做大了" / 运气占成功的 90% |
| 老黄(独立产品) | "批发 API 再零售 / 1 个人 + 小钱起" | Prompt House Mac App $9.98 / Stripe 直接收钱 / 不融资 |
| 深思圈(VC 视角) | 关注融资本身的合理性 | a16z $46B 主权基金 / "造王策略 27-60 天估值翻倍" / Sequoia Agent Economy 三阶段 / 25× 收入倍数 |
真分歧:这不是哲学之争,是定位之争:
- 创新型 + 想做大公司(L4+)+ 有团队 → 看 yitang 7 层光谱 + 深思圈 a16z 视角
- 小微 team / AI 应用 → 王凯反融资路线
- 1 人 + 独立产品 → 老黄路线
- 关注 AI 产业格局 → 深思圈 VC 视角
2.2 分歧 2:科学派(假设驱动) vs 经验派(直觉/勇气驱动)
| 视角 | 立场 | 表述 |
|---|---|---|
| yitang Truman(科学派极端) | "用假设驱动业务 / 没有关键假设的五步法就是个空壳" | Ω 模型 / 5 问灵魂拷问 / "假设"替代"我认为" |
| 王慧文 / 王兴(经验派) | "勇气(王兴答中移动:核心竞争力是什么)" + "运气占成功的 90%" | "校内网逆主流做熟人社交,凭'用户半月访问一次'的观察 + 直觉判断" |
| 王凯(实战派) | "做生意算账 + 多项目并行 + 趁红利期" | 6 心法实战为主,假设思维隐含但不强调 |
| 老黄("Just Do It") | "JUST DO IT,不是等学完才动手" | "无论是做内容还是搞开发,绝对不是等到全部学好了才动手。一切都在前进中解决" |
| 小红书 vibecoding | "审美第一,模型不重要" | 直觉驱动 + 复刻 + 调教 |
真分歧:这是创业哲学的根本分裂:
- 第一次创业 / 风险厌恶高 / 钱不够 → Truman 科学派(系统降低风险)
- 经验丰富 / 已有方向感 / 直觉强 → 王慧文经验派
- AI 工具栈快速迭代期 → 老黄 Just Do It(执行成本极低,假设成本相对升高)
关键洞察(Truman 自己回应):P 型 vs L 型差 100 倍;但 L 型不是消灭 P 型的激情,而是给 P 型 + 调研 + 减法。
2.3 分歧 3:选项池的"加法"做到多大?
| 视角 | 加法目标 |
|---|---|
| yitang Truman(一堂机会预判方法论 §2) | "一般业务至少 10 个选项 / 训练态 MBA 营目标 100 个选项" |
| 王凯(亲历产品列表) | 11 个亲历项目 + 教训 / 不强调"加法"反而强调"快速止损" |
| 王慧文(规模效应) | 不直接给数字 / 强调"在巨大市场找暂时蓝海" |
| 老黄(独立产品时间线) | 6+ 产品并行(newtype.pro / Prompt House / LearnFromYouTube / Super Analyst / Profile / CLI / OS)/ 每个迭代基础上下一个 |
真分歧:加法的目标不同:
- Truman:单一业务方向做加法(找最佳切入点)
- 王凯:多个业务并行做加法(找最佳赚钱机会)
- 老黄:每个产品都迭代加新(找最佳产品形态)
2.4 分歧 4:国内 vs 海外
| 视角 | 立场 |
|---|---|
| 王凯 | "国内卷,海外是必须要去的高地" —— 国内一日比一日卷 / 海外 PMF 后订阅率 3-5 倍 / 单用户 ARPU 高 2-3 倍 |
| yitang Truman | 国内为主(MBA 学员 + 私董会大多国内案例 / 奶茶店青岛 / 一堂自家业务) |
| 王慧文 | 国内为主(校内网 / 美团) / 但提"Uber 输给中国/俄罗斯/印尼,这些地方人口基数大 + 教育水平高,激烈竞争反而提高中国企业国际竞争力" |
| 老黄 | 海外独立产品(北美用户 / Stripe / 报税合规) |
| 深思圈 | 关注美国 AI 行业(a16z / 红杉 / OpenAI / Anthropic / Cursor) |
| 老 D | 国内适配(DeepSeek + 抖音 + 小红书) |
真分歧:这是地理位置 + 用户群 + 产品类型的综合判断:
- AI 应用层 + 卷得无法生存 → 王凯出海路线
- 传统行业(餐饮 / 教育 / 制造)→ yitang / 王慧文国内路线
- 1 人非技术独立产品 → 老黄北美路线
- 关注美国 AI 产业 → 深思圈
3. yitang Truman 派核心(最完整方法论)
3.1 76% 概率数学(最有说服力的一条)
拍脑袋选方向 → 一次成功率 10%
↓
调研后排 90% 不靠谱 → 一次成功率 30%
↓
一年测 4 次 → 总成功率 P = 1 - 0.7⁴ = 76%
把小概率事件变成大概率能跑通。
3.2 15 字心法 + 9 字心态
15 字心法:主动想终局,充分做加法,专业做减法。
9 字底层心态(对应天时地利人和):
| 维度 | 客观面 | 心态 |
|---|---|---|
| 天时(行业机会) | 不取决于你,如行业周期 / 资本风口 | 别拧巴(顺势而为) |
| 地利(商业模式) | 客观规律,如单元模型 / 天花板 | 别天真(快速验证) |
| 人和(初心/团队/资源) | 可以掌控,如团队能力 / 合作 | 别害怕(积极改变) |
3.3 7 层终局光谱图
graph LR
L1["1 副业<br/>极低风险<br/>零花钱"] --> L2["2 个体<br/>低风险<br/>养活自己"]
L2 --> L3["3 单体<br/>中风险<br/>几十-几百万"]
L3 --> L4["4 小规模<br/>中高风险<br/>百万级"]
L4 --> L5["5 体系增长<br/>高风险<br/>千万级"]
L5 --> L6["6 细分头部<br/>极高风险<br/>亿级"]
L6 --> L7["7 做大上市<br/>极高风险<br/>十亿级+"]
classDef low fill:#e8f5e9,stroke:#43a047
classDef mid fill:#fffde7,stroke:#fbc02d
classDef high fill:#ffebee,stroke:#e53935
class L1,L2 low
class L3,L4 mid
class L5,L6,L7 high
规则:选区间不是点 — "保 A 争 B"(如保 3 争 5)。
3.4 五步法 100 版迭代路径
| 版本 | 主题 | 重点 |
|---|---|---|
| V1-V10 | 起盘探索 | 大胆假设 + 快速调研验证 |
| V10-V30 | 内核打磨 | 找到精准产品内核 + 单元模型成立 |
| V30-V50 | 商业模式 | 单元模型基准值 + 动态预测 |
| V50-V100 | 增长 + 壁垒 | 增长引擎 + 长期壁垒构建 |
详见 一堂 / 五步法完整体系(关键假设 + 业务里程碑 + 五步法 5 进阶 + 落地) / 精益创业 / MVP / 关键假设(跨作者主题综合) / 机会预判 / 商业预判 / 终局视角(跨作者主题综合)。
4. 王慧文清华产品课派核心(经典遗产)
4.1 必读 5 本书(王慧文亲选)
| 书 | 价值 |
|---|---|
| The Paypal Wars(支付战争) | 作者真在里面工作过 / Elon Musk 在书里是负面形象 |
| 引爆流行 | 例子鲜活,讲产品的特征 + 迭代为什么被消费者接受 |
| 精益创业 | 把验证速度搞快、成本搞低 / 迭代周期比别人短 = 巨大竞争优势 |
| 创新者的窘境 | 硬盘行业被颠覆的公司不是因为技术,而是组织决定 |
| 零售的哲学(铃木敏文) | 从入门到大成都可以读 / 讲 711 决策过程 |
不推荐:《九败一胜》(美团传记)+ 《腾讯传》(都是宣传用的)。
4.2 5 个 PM 大神 + 王兴金句
| PM | 价值 |
|---|---|
| Neil McElroy(宝洁,经济学专业) | 人类历史第一个产品经理 / 后任宝洁总裁 + 美国国防部部长 |
| 乔布斯 | PC + 图形界面 + iPhone(创新最前沿要学最前沿理论) |
| 张小龙 | Foxmail + QQ 邮箱 + 微信(工程师出身的国内产品大神) |
| Pichai | Chrome(冶金工程专业 / 从 Chrome 业务线晋升而非搜索) |
| Mark Zuckerberg | Facebook(都做过社交网站,认知积累) |
核心:产品经理与专业无关,和本人的三观/价值观/思维方式有关。
王兴答中移动"你们核心竞争力是什么":"勇气"。
4.3 真正的 2 个核心竞争力
- 发现机会的能力 —— 真正的机会来时容易做出成果
- 持续学习进步的能力 —— 美团从社交到团购到外卖,一点一点把能力培养起来
4.4 规模效应 = 商业世界万有引力
"销售/用户量足够大之后(与规模小竞争对手比)所产生的(i)用户体验优势 以及 (ii)成本优势,具体取决于生意模式。"
详见 单元模型 / 规模经济 / 商业模式(跨作者主题综合) / 一堂 / 王慧文清华产品课 66 页完整笔记 / 一堂 / 产品内核 + 单元模型完整体系(五步法第 2+3 步)。
4.5 高频打低频(进入新赛道核心心法)
何时用:做平台战略 / 跨品类扩张时,先做高频品类积累用户基数,再用高频带低频。
美团的"团购 → 外卖 → 酒旅 → 出行"就是高频打低频的典型路径。
5. 王凯算账派核心(AI 时代小微 team)
5.1 6 创业心法(背 6 句够用)
| # | 心法 | 一句话 |
|---|---|---|
| 1 | 赚钱比融资重要 | 互联网时代极少创始团队真正赚到钱 |
| 2 | 算清三件事 | 流量获取成本 / 商业化效率 / 长期运维成本 |
| 3 | 不挑战巨头主战场 | 大模型升级吃不掉啥创业机会,硬件 + 系统 + 生态结合 AI 才吃 |
| 4 | 找需要付费的人 | 不是找用户,是找愿意付出超过获客成本的人 |
| 5 | 多项目并行 | 国内卷在 ROI 低的机会,AI 红利窗口短 |
| 6 | 出海要本地化 | 启动慢 10 倍但 PMF 后模型好太多 |
5.2 算账模型
ARPU 模型 vs 即时付费模型:
| 模型 | 适用 | 关注 |
|---|---|---|
| ARPU(拉长周期算总价值) | 长留产品,3 年总价值 500 元也敢做 | 留存 / 续费 |
| 即时付费(每个用户必须即时回本) | high-intent 需求(写论文 / AI PPT) | 流量洼地 / 单次回本 |
王凯亏 200 美金 0 用户教训:以为所有高 ARPU 模型都适用即时付费。
单技术成本算账:全栈技术月薪 3 万 → 必须做稍微大点的项目,连续成功 5 个小项目的概率低于成功 1 个稍大项目。
5.3 AI 创业者面临的"四座大山"
graph TD
A[小微 AI 创业者] --> B[硬件层<br/>苹果 / OPPO / 特斯拉]
A --> C[系统层<br/>iOS / Android / Windows]
A --> D[平台/生态层<br/>Chrome / 微信 / Slack / Notion]
A --> E[大模型层<br/>OpenAI / Anthropic / Google]
A --> F[大公司应用层<br/>Salesforce / Adobe / 美图]
实战建议:不做 Siri 路径短的 / 不做手机厂商系统肯定会带的 / 找"硬件 + 系统 + 平台 + 大公司"都不愿碰的细分。
5.4 AI 应用层"三步走"长期框架
graph LR
A[Step 1: 用 AI 红利切入<br/>低门槛 + 快验证] --> B[Step 2: 获取流量 / 规模优势<br/>趁红利期攒护城河]
B --> C[Step 3: 延伸业务壁垒<br/>让 AI 占比降到 30% 以下<br/>构建非 AI 护城河]
反面:单纯接 API + 找一个细分场景解决一个具体问题 = 没有壁垒,下一个版本被吃。
5.5 王凯亲历产品列表(11 个项目 + 教训)
| 项目 | 时间 | 结果 | 关键教训 |
|---|---|---|---|
| enjoy 美食电商 | 2018 前 | 公司经历 | 融资 1 亿美金创始团队没赚到钱 |
| AI 电商导购站 | 2023 | 关 | 早期 token 成本算不平 |
| FastFacts(SEC 抓取) | 2023 | 关 | 承载金 30 万天花板 / 量化 team 比我们快 5 秒 |
| MJ for 飞书 | 2023 | 开源 | 流量好但不赚钱 |
| AiPathly(AI 求职) | 2024 | 关 | 创新品做 SEO 错配 |
| Quanta Quest(C 端 wiki AI 搜索) | 2024 | 关 | 数据权限被卡 |
| 松鼠快看 | 2024 | 持续运营 | 国内 C 端规模有限但留存好 |
| AI 戒烟产品 | 2025-2026 | 持续运营 | "把 AI 看作教练" |
| RSI AI 投资策略 | 2025-2026 | 月度 2-4% 收益 | 见 王凯 RSI AI 投资策略与"好资产 好价格 好波动"框架 |
| KeepRule(投资原则 wiki) | 2026 | Demo 上线 | AI Coding 自跑数据库 |
6. 老黄独立产品派核心(非技术 1 人路线)
newtype · Vibe Coding 独立产品实战 §1 Unit Economics 非技术独立开发 3 心法:
| 思路 | 内容 |
|---|---|
| 掂量能耐和资源 | 一个人 + 非技术 + 花点小钱 → 不碰大型/复杂项目 |
| 从商业角度切入 | 一个项目成功不一定要技术壁垒。能不能把技术问题转成商业问题/算数问题?批发 API 再零售 |
| 从强项出发 | 营销、运营、产品是强项,就以这些为骨架架构整个项目,技术作为执行 |
Why 这是非技术独立开发的护城河:
- 程序员擅长的"用最优雅的代码解决问题"在独立开发场景被高估
- 客户不会因为代码漂亮买单,会因为"需求点打得准 + 价格合理 + 体验流畅"买单
- 非技术者的限制反而逼出"先想清楚商业,再用 AI 写代码"的纪律
老黄 6 心法(借鉴 newtype 心法 · 超级个体方法论):
| 心法 | 内容 |
|---|---|
| Unfair Advantage 三能力 | "我有别人无法复制的什么?" |
| Just Do It | 不是等学完才动手 |
| 挖宝藏 vs 挑毛病 | 看待事物两种基本模式 |
| Go Fundamental | 回到基础层(Keynote 替 v0.dev 设计) |
| How vs What | 重 How 不重 What |
| 5 条核心价值观 | 独立思考 / 长期主义 / 第一性原理 / 持续学习 / 创造价值 |
7. 深思圈 VC 派核心(659 篇 a16z / 红杉视角)
入口:深思圈作者全集入口 —— 中文自媒体里"VC + 公司动作"最深的视角
主题 wiki(2026-05-18 补建):
- 深思圈 / AI 创业 0-1 案例 + 增长打法 —— PMF + 0-1 增长 + GTM 完整路径(102+ 篇创业 / 增长案例)
- 深思圈 / AI 一级市场 VC 视角(a16z + 红杉 + YC) —— a16z + 红杉 + YC 三家 VC 视角
- 深思圈 / AI Coding 公司案例(Cursor/Lovable/Bolt/Devin/Replit) —— Cursor / Lovable / Readdy.ai 创业案例
- 深思圈 / 中国华人创业者 + 出海(74 篇) —— 华人 + 出海创业者(74 篇)
7.1 a16z PMF 三条路(Jason Rosenthal 2026-05)
"更多的钱,只是把一个糟糕的结局拖得更久。PMF = 一批真实用户持续地用起来。"
| 路径 | 适用 | 关键 |
|---|---|---|
| 1. 和顶级客户一起做 | 企业级 / 高合规 | 找最挑剔的客户 / 客户需求 = 你的 spec |
| 2. 站在指数曲线前面 | 新基础设施 / 新协议 | 押注下一波技术红利 |
| 3. 自己做自己的第一个客户 | 工程化产品 | dogfood / Cursor / Anthropic 自用 |
7.2 红杉美国 Terrifying Questions 4 问
| # | 问题 | 积极信号 |
|---|---|---|
| 1 | 我的公司有什么存在的权利? | Founder-market fit / 愿意投 10 年 |
| 2 | 人们真的足够在乎吗? | 冷启动 > 关系客户 / 客户眼神发亮 |
| 3 | 我的产品真的能改变行为吗? | "灯泡时刻" / 粘性功能点 |
| 4 | (详见 深思圈 / AI 一级市场 VC 视角(a16z + 红杉 + YC) §2.1) | — |
7.3 标志增长案例(深思圈拆解)
| 案例 | 数据 |
|---|---|
| Lovable | 6 个月 5000 万 ARR / 10 个月百万用户 / 6 招反传统增长 |
| Perplexity | 三年半 4.5 亿 ARR / 5 大增长引擎(被早期投资人泼冷水) |
| Readdy.ai(华人) | 4 个月接近 500 万 ARR / 蓝湖+MasterGo 团队 |
| 9 个月 0 → 月入 10 万美金 | 一个人 / AI 视频 |
7.4 Founder Mode 升级"设计引擎"
"Founder Mode 已经不够用了 —— AI 时代,创始人要从'做引擎'升级到'设计引擎'"(AppSumo 前 CEO)
8. 怎么选 —— 按身份的方法论路径
| 你是谁 | 主用方法 | 起点 |
|---|---|---|
| 学生 / 第一次创业 / 风险厌恶 | yitang Truman 科学派 | 15 字心法 + 76% 概率数学 + 7 层光谱图 → 选项池 100 个 |
| 大厂离职 / 想做大公司 / 有团队 | 王慧文清华产品课派 | 5 必读书 + 5 PM 大神 + 规模效应 ABC + 高频打低频 |
| 小微 AI team / 现金流生存 | 王凯算账派 | 6 心法 + 算账三件事 + 四座大山 + 多项目并行 |
| 1 人 + 非技术 + 独立产品 | 老黄路线 | Unit Economics 3 心法 + 把技术问题转商业 + 强项做骨架 |
| 关注美国 AI 产业 / VC 视角 | 深思圈 VC 派 | 直接 grep ~/kb/sources/learning/深思圈/ |
| 国内不科学上网普通人 + 自媒体变现 | 老 D 13 节课(AI 自媒体与内容工程化(跨作者主题综合)) | 抖音 / 小红书 + DeepSeek + Openrouter + 数字人 30 国语言 |
通用最小步骤(任何身份):
- 想终局(yitang 7 层光谱):选区间不是点 / 保 A 争 B
- 做加法(选项池):至少 10 个,训练态 100 个
- 专业减法:商业模式 / 团队 / 行业 三大硬伤维度
- 算账(王凯三件事 + 老黄 Unit Economics):流量获取成本 / 商业化效率 / 长期运维成本
- 写关键假设清单(Truman Ω 模型):提假设 → 验证 → 迭代
- 快速验证(王慧文 + 王凯 + 老黄):迭代周期比别人短 = 巨大竞争优势
- 多项目并行(王凯):做完赚钱低成本维护,再做下一个
9. 引用清单
主要原始 author wiki:
- 一堂机会预判方法论 —— 15 字心法 / 76% 数学 / 7 层光谱 / 9 字心态 / P型 vs L型 / 4 阶交付物
- 一堂 / 五步法完整体系(关键假设 + 业务里程碑 + 五步法 5 进阶 + 落地) —— 五步法 11 万人上 / 关键假设 Ω 模型 / 1-100 版迭代 / L1-L5 段位
- 一堂 / 王慧文清华产品课 66 页完整笔记 —— 王慧文清华产品课 66 页 / 5 PM 大神 / 5 必读书 / 规模效应 / 王兴金句
- 一堂 / 商业预判进阶(行业画布 + 集中度 + 案例) —— 商业预判进阶 / 行业画布 5 维 / 周子敬集中度 CRn / 4 种市场演化
- 王凯 AI 时代创业方法论与商业判断框架 —— 王凯 516 帖 / 6 心法 / 算账三件事 / 四座大山 / 11 个亲历产品列表
- 王凯 AI 行业判断与大公司竞争格局观察 —— AI 行业判断 / 公司竞争格局 / 三个 Scaling Law 之争
- newtype · Vibe Coding 独立产品实战 —— 老黄独立产品 / Unit Economics 3 心法 / Web → Mac App / Go Fundamental
- newtype 心法 · 超级个体方法论 —— 老黄认知与心法 15 条 / Unfair Advantage / Just Do It / 5 价值观
- 深思圈作者全集入口 —— 深思圈 659 篇 / VC 视角 / a16z / 红杉 / 公司案例库入口
- dontbesilent AI 自媒体课程 / 13 节系统化方法论 —— 老 D 国内 13 节课 / 普通人变现
相关主题 wiki(本系列):
- 机会预判 / 商业预判 / 终局视角(跨作者主题综合) —— 机会预判深度展开(Batch 7)
- 单元模型 / 规模经济 / 商业模式(跨作者主题综合) —— 单元模型 + 规模经济 + 商业模式(Batch 6)
- ROI / 决策评估系统(跨作者主题综合) —— ROI 决策系统(Batch 6)
- 精益创业 / MVP / 关键假设(跨作者主题综合) —— 精益创业 / MVP / 关键假设(Batch 7)
- AI 投资与估值(跨作者主题综合) —— AI 投资估值(Batch 4)
- AI 行业判断与公司竞争(跨作者主题综合) —— AI 行业判断与公司竞争(Batch 2)
History
- 2026-05-18:Topic-dimension wiki 首发。从 10 份 author wiki(yitang 机会预判 / 五步法 / 王慧文产品课 / 王凯创业方法论 / 王凯行业判断 / 老黄独立产品 / 老黄心法 / 深思圈入口 / 老 D 自媒体 / yitang 商业预判进阶)综合改写。显式标注 4 条分歧:融资态度 / 科学 vs 经验 / 选项池规模 / 国内 vs 海外
来源与关联资料
- 一堂机会预判方法论
- 一堂 / 五步法完整体系(关键假设 + 业务里程碑 + 五步法 5 进阶 + 落地)
- 一堂 / 王慧文清华产品课 66 页完整笔记
- 一堂 / 商业预判进阶(行业画布 + 集中度 + 案例)
- 王凯 AI 时代创业方法论与商业判断框架
- 王凯 AI 行业判断与大公司竞争格局观察
- newtype · Vibe Coding 独立产品实战
- newtype 心法 · 超级个体方法论
- 深思圈作者全集入口
- 深思圈 / AI 创业 0-1 案例 + 增长打法
- 深思圈 / AI 一级市场 VC 视角(a16z + 红杉 + YC)
- dontbesilent AI 自媒体课程 / 13 节系统化方法论
- 机会预判 / 商业预判 / 终局视角(跨作者主题综合)
- 单元模型 / 规模经济 / 商业模式(跨作者主题综合)
- ROI / 决策评估系统(跨作者主题综合)
- 精益创业 / MVP / 关键假设(跨作者主题综合)
- AI 投资与估值(跨作者主题综合)
10. 42章经跨期增补：让方向、组织与商业证据互相约束
2026-10-18跨期整理增补。以下综合2023—2026年访谈，不把后来的认识倒填到早期，也不把融资、估值、榜单或原型数量等同于经营成功。各期自述数字未审计；推演和可迁移方法属于编辑综合，具体出处链接到节目原文。
10.1 选择重要问题，再盘可控制的条件
Albert在2024年PMF访谈中建议寻找用户预期清楚、当前能力能交付、存在多个玩家且可研究后来者路径的市场。他说“做没壁垒的事”，并非永远不积累优势，而是不要把创业第一步押在只有自己想象、没有需求或可实现性证据的事情上。外部条件包括用户任务与替代品、分发、付费习惯；内部条件包括实际能力、现金和可承受试错。能估成功概率与能承受损失是两回事，主观把握不能被包装成精算。节目原文
到2026年3月，Albert进一步把“Why选重要问题”与“How把事情做好”分开。他没有主张只能做熟悉的小事：可以进攻新方向，但先找能控制的变量，逐步建立优势，而非靠竞争者恰好失败。曲凯更重视抢先封装技术、获取数据和资本资源；Albert追问“赢”究竟指融资、估值、赚钱，还是用户持续需要。分歧没有被消除：抢先可以改善下一步条件，却不能自动证明已经创造可持续价值。对已有盈利业务，停止追逐每个技术热点也可能合理；旧组织不能支持新机制时，另建小闭环比强改整家公司更可行。节目原文
实际立项可写三层：第一，什么问题值得解决、用户不用我们会怎么办；第二，我们已控制了什么、还赌了什么；第三，如果赌错，何时会看到、能承受多大代价。不要把“方向比执行重要”读成不用交付，也不要把“快速试错”读成省略问题判断。一次实验发现归因错误并修正，也属于有效正反馈，不必靠漂亮结果维持士气。
10.2 同一个人的观点会变，产品阶段也会变
Albert的跨期变化尤其值得保留。2024年他重视明确交付、付费和可持续资产，放弃一次性热潮式工具；2026年3月又反思80分AI娱乐内容面对90分成熟内容的吸引力，模型新颖不应成为用户支付“AI税”的理由。到2026年5月，其团队两个月做出几十个内部产品，却承认当时收益为零、尚未达到对外产品标准；方向转向makers之间的“回响”，即被在意的人理解、合作和二次创造。节目原文 节目原文 节目原文
这不是把所有观点缝成一句“始终坚持PMF”。后一期更重视影响、体验与连接，但他同时区分个人免费创作与有工资、股东责任的公司。回应是被看到，回响是被相关人理解，回报是经济收入；三者可相互促进，不能互相替代。对创作者平台，原型数量、点赞数与收入应分别记录；对陪伴产品，被理解的主观体验要与复访、长期满意和安全反馈交叉检验。产品类型不同，商业验证链可以不同，但公司必须知道自己在哪一环尚未闭合。
陈昱2023年的技术投资经历提供另一种连续性：Zilliz从GPU数据库走向向量数据库时，没有把旧积累全部扔掉，而是保留计算、数据库与实现能力，用客户以图搜图的新需求确认转向；原型容易不代表性能和全球商业化容易。可迁移的方法是列清转向中保留的能力、丢弃的假设以及新增客户证据，而不是为了追新词从零重启。他也承认未预测到ChatGPT后来的影响，专业理解增加判断条件，不赋予预知未来的能力。节目原文
梦琪2026年反省提供了更具体的反例：因为通用Agent拥挤，先选“找人”垂直赛道，再把达人、招聘、专家、客户视为同一能力外延，最终发现共同检索技术不等于同一使用者任务。专业营销人员要过程可控，创始人要直接拿结果；未选定目标角色时，听取每条反馈可能把产品拖向彼此冲突的形态。转型要记录旧假设被什么事实否定、新扩展补的是哪段交付，而不是只扩大融资市场叙事。节目原文
她也反省因自我认同和融资故事提前招算法人才、讲强化学习，后来将精力转向用户意图、上下文切换、稳定性和问题验收。对应的纪律不是“永不招算法”，而是每笔技术投入都对应当前可测瓶颈。匿名同行指控不作为行业事实；内部每天高密度使用的Skill比旧产品更有效，也只说明在数据、权限、目标一致的环境下取得进展，外售还要补客户差异、授权和维护。
10.3 时机由多种时钟构成，不由新闻热度决定
周健2024年观察到模型、预算、甲方关注、算力采购和应用供给不同步。他认为机会来自这些变量交汇，又明确区分客户请你免费培训与已经下决心试用。甲方先买GPU再找应用，可能给算力利用率管理创造暂时窗口；他本人就提醒这类机会未必持久。不能把沉没支出直接认作持续需求，也不能将2024“应用会爆发”的预测当作后续实际增长证明。节目原文
曲凯在2024年Character.AI单口里强调资本窗口：数据改善可能抵不过估值倍数收缩，融资进度与产品进展是两种时钟。他据交易推演模型商品化及“模型战争近结束”，属于当时观点；单一交易不足以证明行业终局。其“宿命论”提示外部约束，却不应让团队放弃可控行动。真正可做的是用现金跑道、采购周期、用户成熟度和供应商依赖设情景，而不是赌下一轮必能按更高估值融资。节目原文
Jenny Xiao在2025年提醒，漂亮高估值也提高下一轮所需收入门槛；二三家客户贡献多数收入，有时更像高质量定制服务，未必已验证规模化产品。把市场规模、已签约、实际回款、标准模块复用和续费分开看，才能避免用一个好故事覆盖五个尚未验证的问题。节目原文
10.4 组织方法要匹配业务，不照搬胜者外形
魏小康以链条长短、反馈周期与毛利解释组织差异：高毛利、短链条业务较能先找好人才再匹配岗位；低毛利、长链条业务更需要经营计划、资源预算和协作精度。这个坐标不穷尽所有组织，却能帮助判断该学什么。三人也可能理解不同，应对齐目标、责任和滚动计划；不需要为此先复制大厂层级、职级和长篇绩效材料。节目原文
人才密度应指匹配当前任务、能协同解决难题的人，而非名校、工资或年龄密度。招聘先定业务难题、何时需要谁，再建设供给与触达；一次面试之外，还要看真实工作证据与经允许的可靠背调。Offer需讲清双方预期，不以宏大意义替代公平报酬。魏对年轻创始人及OPC较悲观，曲凯则指出老手可能受规则束缚、年轻人可能打开新路；把分歧转换成可验证的资源、组队和任务能力，比按出生年份选人更可靠。
周健从十人到千人组织的经历补充“起飞”后的刹车：早期成员会遇到成长上限，外来经理也可能不适应零到一；新旧成员、增长节奏、现金与责任都需重新匹配。资历不等于忠诚，老兄弟也不天然胜任。组织目的是交付，不是批量淘汰；真实任务中的成长需要支持和合适战场，不应变成“招进来筛掉”的借口。
Palona任川的2025年实践，则把AI团队按最终商家/消费者结果分工，工程直接接触客户、减少销售—PM—工程的信息损耗。约20人的做法不能直接证明50人或100人仍无需产品与治理角色；他说不知道。默认让AI先尝试，也不等于取消责任。其快速审查和回滚经验在问答中受到P0、脏数据和不可逆影响的挑战，后来也承认Infra更保守。可以缩短交接，但生产权限、安全审查和最终验收仍应按风险独立设置。节目原文
范皓宇2025年提供了组织与产品评审的另一面：责任可以端到端，但不能把“完整”理解为每个人无限全能。他用从发现、初用、持续探索到换代迁移的体验路径，提醒团队不要把单一环节改善当完整交付。用户习惯、已有内容和权益也会形成软件的慢变量，快速改版并非无成本。对应组织设计，需有人对整条用户旅程负责，同时保留专业协作和风险门禁。节目原文
他强调品味和整体感受，曲凯追问品味可能错或超前，范也承认要承担结果。不能将这段化约成“数据无用”，也不能以数据好消除所有体验反证。更实用的结合是：将不对劲写成具体假设，查看分群、长期使用、失败案例和真实原话，再调整实现；“既要又要”是寻找降低冲突的设计，不是取消预算与取舍。由此，创业者对完整体验的坚持，应与可被推翻的判断共同存在。
10.5 增长是产品系统的一部分，但冲刺有边界
ACE Studio的Joe在2025年回顾HF0：已经有付费基础的团队，通过12周驻留、单一目标、每周结果反馈改变工作方式。最戏剧性的日收入翻倍来自提价一倍，后续还有价格模型、KOL触达SOP、优胜视频广告放大、模型升级和流失用户唤回。不是信念让现金凭空出现，也不能把最终十倍增长全归一招。其前六周的账户、归因、转化与素材基础设施不能跳过。节目原文
可复用的是把功能、价格、触达和服务放进同一商业循环，逐个检验。提价短期收入上升还要看后续退款、续费和分群留存；广告ROI自述没有完整毛利和归因口径，不能直接作为扩量参数。单一冲刺指标要配质量与安全护栏。离营后团队没有保持极端工时，Joe转而按周期分别做研发和增长冲刺；有限高强度不等于永久工作制度。
信念在这里的合理位置，是让团队愿意探索更多候选方案、忍受反馈迟滞，而非停止承认坏消息。张津剑在2024年生命力访谈强调稳定注意力、困难选择和不把挫折变成自我否定；曲凯同时认为信念需要研究支撑，两人对“信念需不需要理由”存在真实差异。有关频率、能量和物理的表达只是类比，不能当作心理学因果或成功科学。对于商业决策，可把对自身价值的肯定与对业务假设的否决并存：不必否定自己，也必须允许项目被证伪。节目原文
10.6 资源型打法先问条件，资本叙事和实业结果分开验收
资源不是原罪，也不是每个团队都能复制的路线。周健的央国企交付依靠既有行业人才、商业资源与长销售链；大客户算力帮助产品锤炼，但能否形成高毛利标准模块尚需后续证明。魏小康讨论的高薪抢人，要有业务回报和现金支撑；Albert批评用字节式大量顶尖人才和资本赛马作为普通团队默认方案，指出良好生意与文化也可以让普通成员持续做对事。先认清自己是个人收入、可持续小公司还是风险投资型扩张，再选择打法。
每个项目至少分别保留四类证据：用户得到什么、企业交付后留下什么、钱怎样回到经营主体、资本方为何愿意定价。融资可能改善供给和生存条件，却不能代替续费；退出可能让部分股东获得回报，却不能证明原产品盈利；创始人的技术愿望也不能替代用户任务。Character.AI材料里的创始人心理和交易动机尤其应保留“曲凯推论”的归属，不能以事后结果包装成必然命运。
给产品团队的最小落地表，是每月重看一页：当前最重要用户问题、三项可控条件、三项外部假设、下一次实证、现金与时间上限、谁对结果负责、什么情况下停止。团队规模与工具采用只是手段；如果用户价值、经营结果和个人目标相互冲突，要把冲突写出来做取舍，而不是用更大的故事盖过去。
English version
10. A Synthesis Across 42 Zhangjing Episodes: Let Direction, Organization, and Business Evidence Constrain One Another
Added through a synthesis across episodes on 2026-10-18. The following draws on interviews from 2023–2026. It does not project later insights backward into earlier periods or equate funding, valuations, rankings, or prototype counts with business success. Self-reported figures from individual episodes are unaudited. Inferences and transferable methods are editorial synthesis; the source links point to the original episodes.
10.1 Choose an Important Problem, Then Assess the Conditions You Can Control
In his 2024 PMF interview, Albert recommended looking for markets where users have clear expectations, current capabilities can deliver, multiple players exist, and the paths taken by later entrants can be studied. His advice to “do things without barriers to entry” did not mean never building an advantage. It meant avoiding a first entrepreneurial step staked on something that exists only in one's imagination, without evidence of demand or feasibility. External conditions include user tasks and alternatives, distribution, and payment habits; internal conditions include actual capabilities, cash, and affordable experimentation. Estimating the probability of success and being able to absorb a loss are different matters. Subjective confidence must not be presented as actuarial precision. Original episode
By March 2026, Albert further separated “why: choose an important problem” from “how: do the work well.” He was not arguing that founders should pursue only small, familiar tasks. A team can enter a new direction, but should first identify controllable variables and gradually build an advantage, rather than rely on competitors happening to fail. Qu Kai placed more emphasis on being early to package technology and obtain data and capital resources. Albert asked what “winning” actually meant: funding, valuation, making money, or continuing user need. The disagreement remained unresolved: moving first can improve the conditions for the next step, but does not automatically demonstrate sustainable value. For an already profitable business, declining to chase every new technology trend may also be reasonable. When an existing organization cannot support a new mechanism, building a separate small, complete operating loop may be more feasible than forcing the entire company to change. Original episode
An actual project proposal can be written in three layers. First, what problem is worth solving, and what would users do without us? Second, what do we already control, and what are we still betting on? Third, if the bet is wrong, when will we know, and how much can we afford to lose? “Direction matters more than execution” does not excuse a failure to deliver; “experiment quickly” does not mean skipping judgment about the problem. An experiment that exposes and corrects faulty attribution also provides useful positive feedback. Morale need not depend on attractive results.
10.2 A Person's Views Can Change, and So Can a Product's Stage
Albert's changes across episodes are particularly worth preserving. In 2024, he emphasized clear delivery, payment, and sustainable assets, moving away from tools built around one-off waves of excitement. In March 2026, he questioned the appeal of AI entertainment content scoring 80 when mature content scored 90: model novelty should not make users pay an “AI tax.” By May 2026, his team had built dozens of internal products in two months, yet acknowledged that returns were zero at that point and that the products had not reached the standard for external release. The direction shifted toward “resonance” among makers: being understood by people who matter to them, collaborating, and creating further work from one another's creations. Original episode Original episode Original episode
These views should not be stitched together into “he consistently pursued PMF.” The later episode gave greater weight to impact, experience, and connection, while still distinguishing free personal creation from a company with salaries and obligations to shareholders. Response means being seen; resonance means being understood by relevant people; return means economic income. The three can reinforce one another, but cannot replace one another. A creator platform should record prototype counts, likes, and revenue separately. A companion product should cross-check the subjective experience of being understood against returning use, long-term satisfaction, and safety feedback. Different product types may have different chains of commercial validation, but a company must know which link remains incomplete.
Chen Yu's account of technology investment in 2023 offers a different form of continuity. As Zilliz moved from a GPU database toward a vector database, it did not discard all its earlier work. It retained capabilities in computation, databases, and implementation, and used new customer demand for image-to-image search to validate the pivot. An easy prototype did not imply easy performance engineering or global commercialization. The transferable method is to list the capabilities preserved in a pivot, the assumptions discarded, and the new customer evidence gained, rather than restarting from zero to chase a new term. He also acknowledged not predicting ChatGPT's subsequent impact. Domain expertise adds conditions for judgment; it does not grant foresight. Original episode
Mengqi's reflections in 2026 offer a more specific counterexample. Because general-purpose agents were crowded, the team first chose the vertical of “finding people,” then treated influencers, recruitment, experts, and customers as extensions of the same capability. It ultimately found that shared retrieval technology does not mean a shared user task. Professional marketers want control over the process; founders want the result delivered directly. Without choosing a target role, listening to every piece of feedback can pull a product into conflicting forms. A pivot should record which facts contradicted the old assumptions and which part of delivery the new expansion addresses, rather than merely enlarging the market narrative used for fundraising. Original episode
She also reflected on hiring algorithm specialists and discussing reinforcement learning prematurely because of self-image and the fundraising story, then later shifting effort to user intent, context switching, stability, and acceptance of completed work. The corresponding discipline is not “never hire algorithm specialists,” but to connect every technical investment to a currently measurable bottleneck. Anonymous accusations against peers are not treated as industry facts. That a Skill used intensively inside the company every day worked better than the old product only demonstrates progress in an environment with aligned data, permissions, and goals. External sales would still require addressing customer differences, authorization, and maintenance.
10.3 Timing Is Made Up of Several Clocks, Not News Attention
In 2024, Zhou Jian observed that models, budgets, customer attention, compute procurement, and application supply were out of sync. He saw opportunities at the intersection of those variables, while explicitly distinguishing a customer inviting you to give free training from one that had decided to run a trial. Customers buying GPUs before finding applications could create a temporary opening for compute-utilization management; he himself warned that such opportunities might not last. Sunk expenditure cannot simply be treated as continuing demand, nor can a 2024 prediction that “applications will take off” serve as evidence of growth that actually followed. Original episode
In his 2024 solo episode on Character.AI, Qu Kai emphasized the capital window: improvements in operating metrics might not offset shrinking valuation multiples. Fundraising progress and product progress run on different clocks. His transaction-based inferences about model commoditization and a “model war nearing its end” were views held at that time; a single transaction cannot establish an industry's final state. His “fatalism” highlights external constraints, but should not make teams abandon controllable actions. The actionable response is to build scenarios around cash runway, procurement cycles, user readiness, and supplier dependence, rather than betting that the next funding round will necessarily come at a higher valuation. Original episode
In 2025, Jenny Xiao warned that an attractive, high valuation also raises the revenue threshold needed for the next round. When two or three customers contribute most of the revenue, the business may resemble high-quality custom services rather than an already validated scalable product. Market size, signed contracts, cash actually collected, reuse of standardized modules, and renewals must be assessed separately to avoid covering five untested questions with one good story. Original episode
10.4 Match Organizational Methods to the Business, Rather Than Copying Winners' Outward Form
Wei Xiaokang explained organizational differences through value-chain length, feedback cycles, and gross margins. A high-margin business with a short chain is better positioned to hire strong people first and then match them to roles; a low-margin business with a long chain needs greater precision in operating plans, resource budgets, and coordination. These dimensions do not describe every organization, but can help determine what to learn from one. Even three people may understand a task differently; align goals, responsibilities, and rolling plans without first copying large-company hierarchies, job grades, and lengthy performance documents. Original episode
Talent density should mean people suited to current tasks who can solve difficult problems together, rather than a concentration of elite-school credentials, high salaries, or a particular age. Hiring starts by specifying the business problem and who is needed when, then building the candidate pool and outreach. Beyond one interview, examine evidence of real work and reliable background checks conducted with permission. An offer should clarify expectations on both sides; a grand sense of purpose does not replace fair compensation. Wei was relatively pessimistic about young founders and one-person companies (OPCs), while Qu Kai noted that veterans may be constrained by established rules and young people may find new paths. Translating this disagreement into testable resource, team-building, and task capabilities is more reliable than selecting people by birth year.
Zhou Jian's experience with organizations growing from ten people to a thousand adds the need for brakes after “takeoff.” Early employees may reach growth limits, while outside managers may struggle with zero-to-one work. New and existing members, growth pace, cash, and responsibilities all need to be matched again. Seniority does not equal loyalty, and longtime companions are not automatically qualified. An organization exists to deliver, not to eliminate people in batches. Growth through real work requires support and an appropriate setting; it should not become an excuse to “hire them and screen them out.”
In Palona's 2025 work, Ren Chuan organized AI teams around final outcomes for merchants and consumers. Engineers dealt directly with customers to reduce information loss across sales, PM, and engineering handoffs. What worked at approximately twenty people does not demonstrate that a team of fifty or a hundred still needs no product or governance roles; he said he did not know. Letting AI try first by default does not remove accountability. During the Q&A, his experience with rapid review and rollback was challenged over P0 incidents, dirty data, and irreversible effects. He later acknowledged a more conservative approach to infrastructure. Handoffs can be shortened, but production permissions, security review, and final acceptance should still be set independently according to risk. Original episode
In 2025, Fan Haoyu offered another side of organizational and product review. Responsibility can be end to end, but “complete” should not mean every individual is infinitely versatile. His experience path—from discovery and first use, through continued exploration, to migration across product generations—reminds teams not to mistake an improvement in one step for complete delivery. User habits, existing content, and entitlements also create slow-moving variables in software; rapid redesign is not cost-free. The corresponding organizational design needs someone accountable for the entire user journey while retaining specialist collaboration and risk gates. Original episode
He emphasized taste and the overall experience. Qu Kai asked whether taste might be wrong or ahead of its time, and Fan acknowledged the obligation to bear the consequences. This exchange cannot be reduced to “data is useless,” nor can good data erase all contrary evidence about experience. A more useful combination is to turn the feeling that something is wrong into a specific hypothesis, examine cohorts, long-term use, failure cases, and actual user quotations, then adjust the implementation. Wanting both sides of a tradeoff means seeking designs that reduce the conflict, not removing budgets and choices. A founder's commitment to the complete experience should therefore coexist with judgments that can be overturned.
10.5 Growth Is Part of the Product System, but Sprints Have Limits
In 2025, ACE Studio's Joe looked back on HF0: a team with an existing base of paying customers changed how it worked through a twelve-week residency, a single goal, and weekly outcome feedback. The most dramatic doubling of daily revenue came from doubling the price. Later work included pricing models, a standard operating procedure for KOL outreach, scaling winning video advertisements, model upgrades, and bringing back churned users. Belief did not conjure cash from nowhere, and the eventual tenfold growth cannot be attributed entirely to one move. The account, attribution, conversion, and creative infrastructure built during the first six weeks cannot be skipped. Original episode
The reusable approach is to place features, pricing, outreach, and service in one commercial loop and test them individually. A short-term revenue increase after raising prices still requires examining later refunds, renewals, and cohort retention. Self-reported advertising ROI without complete gross-margin and attribution definitions cannot be used directly as a scaling parameter. A single sprint metric needs quality and safety guardrails. After leaving the program, the team did not maintain extreme working hours. Joe instead alternated R&D and growth sprints by period; a limited spell of high intensity is not a permanent working system.
Belief has a reasonable role here: encouraging a team to explore more candidate approaches and tolerate delayed feedback, rather than stop acknowledging bad news. In his 2024 interview on vitality, Zhang Jinjian emphasized steady attention, difficult choices, and not turning setbacks into self-rejection. Qu Kai argued that belief needs research behind it; the two genuinely differed over whether belief needs reasons. References to frequency, energy, and physics are analogies, not psychological causation or a science of success. In business decisions, affirming one's own worth can coexist with rejecting a business hypothesis: a project must be allowed to be falsified without requiring a rejection of oneself. Original episode
10.6 Ask What a Resource-Intensive Strategy Requires, and Validate Capital Narratives Separately from Operating Results
Resources are neither inherently wrong nor a path every team can reproduce. Zhou Jian's delivery to central and other state-owned enterprises relied on existing industry talent, business resources, and long sales cycles. Large customers' compute resources helped refine the product, but whether that could produce high-margin standardized modules still needed later evidence. Wei Xiaokang's discussion of competing for people with high salaries requires business returns and cash to support it. Albert criticized treating a ByteDance-style approach—large numbers of top people and capital competing across parallel bets—as the default for ordinary teams. He noted that a sound business and culture can also help ordinary team members consistently do the right things. First determine whether the goal is personal income, a sustainable small company, or venture-backed expansion, then choose the approach.
Every project should retain at least four distinct categories of evidence: what users receive; what assets remain with the company after delivery; how money returns to the operating business; and why capital providers are willing to assign a valuation. Financing can improve supply and survival conditions, but cannot replace renewals. An exit may provide returns to some shareholders without demonstrating that the original product was profitable. A founder's technological ambition cannot replace the user's task. In particular, accounts of founder psychology and transaction motives in the Character.AI material must retain their attribution as Qu Kai's inferences, rather than turning hindsight into an inevitable destiny.
A minimal implementation for a product team is a one-page review each month: the most important current user problem; three controllable conditions; three external assumptions; the next empirical test; cash and time limits; who owns the result; and conditions for stopping. Team size and tool adoption are means. If user value, operating results, and personal goals conflict, write down the conflict and make a choice instead of covering it with a larger story.
2026-10-20 增补：短期新词工具站，怎样把热点变成可交付的订单
绪轩（@xuxuanzhou）在 2026 年 10 月 19 日发布了一次 Bulin 47 热梗视频工具站复盘。产品只做一件事：用户上传照片，生成自己出现在梗视频中的成片。以下是对这一个案例的方法整理；经营数字和时间线均来自作者自述及其附图，未独立审计，不能用来预测其他站点的收入。
先分清流水、标价与利润
作者在开头使用含税 Stripe 总交易额描述累计流水，随后在“数据”小节明确：标题中的 72 小时 $5,480 是首单后 72 小时的订单标价，不含结账加税、未扣退款；第一周 558 单、393 位付款者、$6,434 也采用标价口径，退款 22 单。附图另展示累计总交易额 $7,197.19 和净销售额 $6,062.52，时间范围与前两者不同。原文还提到截至发帖时 API 花费超过人民币一万元。金额不能混算，也不能把“净销售额”当作扣完模型、支付、运维及人力成本后的利润。
找到“想复刻”的任务，再判断窗口
作者先在旧站测试 hotel lobby 页面，再让 Claude Code 协助从 TikTok 寻找相似的新梗，并用 Google Trends 比较候选词。他选择 Bulin 47 的理由是趋势刚起、出现复刻需求、相关域名仍可注册。这里值得复用的是把流行内容与具体工具任务接起来，而不是只追搜索量。趋势曲线、可注册域名与少量成功案例，都不足以单独证明付费需求。
作者报告订单高峰很短：10 月 10 日至 15 日每天分别为 180、185、100、33、24、25 单，并称 85% 收入集中在头 72 小时。这提示短期工具要提前准备登录、支付、积分和埋点，减少从发现机会到交付的时间；不能把这一条曲线推成所有热点的寿命，也不能用“快”跳过付费、生成与下载的基本验收。
收费要匹配使用任务，改价后出单不等于因果成立
作者发现默认推荐月付订阅不适合“做一条视频就走”的用户，于是改为优先展示积分包；其记录显示改动后约八分钟出现首单，并称 $9.99 单条订单占 92%。这支持一个待验证的判断：一次性消费任务可以先比较按次或积分包与订阅的匹配程度。它不是所有 AI 工具都应取消订阅的结论，前后对比也未排除流量、时间和其他页面改动的影响。
AI 写出的业务规则也要审，不只审代码能不能跑
正文和故障记录图显示，AI 曾加入每日限额和“没换人就判失败”的规则，导致正常付费成片被拒或自动退款。其他故障涉及数据库连接、移动端上传被导航遮挡、付款后找不到生成入口、登录前后照片丢失，以及上游成片未成功转存。编辑建议：上线前逐项确认谁批准限额、什么证据算失败、退款如何触发；沿“付款—上传—生成—结果保存—下载”走一次真实流程。保留必要的安全、版权和反欺诈约束，不能把修复误判理解为删除所有保护。
用分段漏斗定位问题，给结论保留边界
作者附图把“打开付款页后付掉的比例”列为 iPhone 54%、Android 17%，另把全体访客的付款比例按来源列为直接访问 5.8%、ChatGPT 3.9%、Google 2.3%。这是两种不同分母，不能相互替代；付款方式、地区和用户意图也未被控制。作者关于 Apple Pay 更顺畅的解释仍是假设，ChatGPT 访客转化较高也不证明 llms.txt 或 IndexNow 带来了增长。原文还记录向 100 名弃单者发送 $6.99 优惠邮件后无人购买，说明一次降价并没有修复这批流失，不能推成召回永远无效。
何时用：面对快速消退、目标动作明确的工具需求时，先记录发现时间、需求证据、交付成本与停止条件，再观察每一步的失败。可复用的是验证和交付过程；盈利能力仍需用同一周期内扣除退款及全部成本后的结果判断。
来源：绪轩的新词站完整复盘（X，2026-10-19）。本文为独立整理；原文及六张配图保留在私人知识库。
October 20, 2026 addition: turn a short-lived search trend into an order you can deliver
On October 19, 2026, Xuxuan (@xuxuanzhou) published an account of a tool built around the Bulin 47 video meme. Its task was narrow: users uploaded a photo to generate a video placing themselves in the meme. This section extracts methods from that single case. Business figures and the timeline come from the author's account and attached images, have not been independently audited, and cannot predict revenue for another site.
Separate transaction volume, order prices, and profit
The opening uses Stripe transaction totals including tax to describe cumulative volume. The later data section clarifies that the headline $5,480 represents order prices during the 72 hours after the first order, excluding checkout tax and before refunds. The first-week figures of 558 orders, 393 paying customers, and $6,434 use the same price basis, with 22 refunded orders. A separate screenshot shows cumulative transaction volume of $7,197.19 and net sales of $6,062.52 over a different reporting window. The author also reports API spending exceeding RMB 10,000 by the time of publication. These figures must not be combined, and net sales are not profit after model, payment, operating, and labor costs.
Identify the task people want to reproduce, then assess the window
The author first tested a hotel lobby page on an existing site, then used Claude Code to help find similar emerging TikTok memes and compared candidate terms in Google Trends. Bulin 47 was selected because interest was beginning to rise, people wanted to reproduce the video, and relevant domains were still available. The useful method is connecting popular content to a specific tool task, rather than chasing search volume alone. A trend curve, an available domain, or a handful of successful examples cannot independently establish willingness to pay.
The reported order window was short: daily orders from October 10 through 15 were 180, 185, 100, 33, 24, and 25. The author states that 85% of revenue arrived in the first 72 hours. This suggests preparing login, payments, credits, and analytics before an opportunity appears to shorten the path to delivery. That one curve does not establish the lifespan of every trend, and speed does not justify skipping basic checks of payment, generation, and download.
Match charging to the task; orders after a pricing change do not establish causation
The author found that a default monthly subscription did not fit users who wanted to make one video and leave, so credit packs became the primary offer. The account places the first order about eight minutes after that change and says $9.99 single-video orders represented 92% of orders. This supports a hypothesis worth testing: for an occasional task, compare per-use pricing or credits with subscriptions. It does not mean every AI tool should remove subscriptions. The before-and-after observation also does not control for traffic, timing, or other page changes.
Review AI-generated business rules as well as working code
The text and incident log describe daily limits and a rule that marked a result as failed when it appeared not to replace the person. These rules rejected normal paid outputs or triggered refunds. Other incidents involved database connections, mobile navigation covering uploads, missing generation entry points after payment, photos lost across login, and outputs not being successfully copied from the upstream service. An editorial application is to confirm who approved each limit, what evidence constitutes failure, and how refunds are triggered. Test the complete payment–upload–generation–saved result–download journey. Necessary safety, copyright, and fraud controls must remain; fixing false positives does not mean removing all protections.
Use separate funnel stages and retain uncertainty
An attached chart reports payment completion after opening checkout at 54% for iPhone and 17% for Android. Another reports payment as a share of all visitors by source: 5.8% for direct visits, 3.9% for ChatGPT, and 2.3% for Google. These have different denominators and are not interchangeable. Payment methods, geography, and intent were not controlled. The proposed Apple Pay explanation remains a hypothesis, and stronger conversion from ChatGPT visitors does not establish that llms.txt or IndexNow caused growth. The author also reports that none of 100 checkout abandoners bought after receiving a $6.99 offer by email. That discount did not recover this group; it does not establish that recovery campaigns never work.
When to use this: for a clearly defined tool task with rapidly declining demand, record when the opportunity was discovered, evidence of demand, delivery costs, and stop conditions, then observe failures at each step. The reusable lesson is the validation and delivery process. Profitability still requires a consistent reporting period with refunds and all costs deducted.
Source: Xuxuan's full account of the emerging-keyword site (X, October 19, 2026). This is an independent synthesis; the original text and six images are retained in the private knowledge base.