中文精读
FlowGPT的早期增长：让创作者完成体验，再判断用户为什么回来
42章经单期精读 · AI创作平台
嘉宾：Jay 党嘉成，当期为FlowGPT创始人｜主持：曲凯
发布：2024年3月14日｜时长：49分57秒｜整理：2026年10月18日
录制：2023年11月；因融资细节延迟发布
官方单集与节目介绍
一个Prompt社区，为什么开始把自己叫应用平台
Prompt是交给模型的指令、背景和规则。分享它可以像分享模板，但使用者真正要的可能是一场文字冒险、一个聊天角色，或一个能反复完成任务的工具。FlowGPT的定位变化，正是从收藏指令转向在平台上直接消费这些体验。
官方介绍称，截至2023年11月平台有300万月活用户，并在本期发布时宣布新一轮千万美元融资。数字是节目及当事人口径，未独立审计，也不代表今天规模。更值得精读的是：平台怎样组织供给、识别复访，以及回应模型厂商也做应用商店的竞争。
1. 两天做出第一版，随后用人工接触补足判断
线索：03:13–08:37。 Jay说第一版花了两天，第二周看到用户在社交平台自发传播，才决定辍学投入。他没有在启动前找到单一、强烈的痛点，而是先把愿意花时间写Prompt的人聚起来。
这不是只有一个灵感的故事。他从高中做项目、读产品与创业材料，在伯克利有大量合作经历；早期团队也来自曾经合作、被证明可靠的人。所谓“没有经验”的优势，是个人牵挂少、愿意快速尝试，并不能推出经验会妨碍创业，更不能把辍学当作成功公式。
真正费力的工作发生在发布之后：他到Discord、Reddit、Twitter寻找创作者，称最初两周私聊五六百人，累计几千人。有人直接批评产品，他仍继续做。人工招募同时补供给与理解：谁愿意认真创作、在哪里卡住、为什么会分享，都比泛泛问“你喜不喜欢AI”具体。
编辑推论：适合迁移的是低成本原型加高密度接触。只有页面浏览、没有作品发布与消费，尚不能证明一个双边平台已经启动；Jay的个案也没有提供人工触达的转化率或可复制获客成本。
2. 多模型开放，必须落到创作者的完整作品上
线索：13:00–16:26、19:35–22:56。 Jay将GPT Store比作iOS，将FlowGPT比作Android，表达的战略是让创作者选择模型，减少被单一模型生态绑定。他的前提是消费者会跟着创作者走。
这个比喻只有落实到体验才有意义。Jay认为模型目录常让用户选好零件后离站部署，而FlowGPT希望把生图等能力嵌入作品，让用户在站内完成消费。例如文字冒险可以在剧情中生成场景图，而非要求玩家先找模型、部署，再返回故事。Bot只是可消费内容的一种形式。
团队试做过很多应用，却认为深度创作者对Prompt和玩法的理解更强，因此愿意把分发交给用户。创作者服务也不仅是模型列表，还包括曝光、反馈被听见和持续维护作品的条件。
节目对Character.AI的负面评价来自Jay所述的创作者访谈与主持人回应，不能写成对竞品运营的客观定论。编辑推论：开放能否形成优势，要看选择模型是否真的改善作品、创作者是否持续更新、消费者是否随之迁移；节目没有给出足以确认网络效应的数据。
3. 用户换了Bot，可能仍在重复消费同一品类
线索：17:04–18:38、20:25–23:54、26:44–28:08。 Jay区分工具与娱乐：工具使用者可能希望持续使用同一个功能；娱乐使用者可能玩三天一个故事，再寻找新故事。Bot在这里承载的是创意，类似视频中的内容单元。
他列出的热门体验包括Prompt生成器、陪伴角色，以及结合生图的文字游戏；称当时约60%的使用与娱乐相关。该比例来自平台观察，节目未交代分母、窗口和分类方法。它不能说明所有AI用户以娱乐为主。
更有用的是留存层级的区分：有人每周回来用同一应用，也有人每周回来寻找同一品类的新内容。后者换作品，并不自动等于平台失去价值。反过来，常回来尝鲜也不证明与某个角色建立了长期关系。
Jay还观察到有人将页面常挂，想找AI时就打开。主持人由此讨论入口竞争；这只是方向信号，缺少常驻比例及持续时间。编辑推论：先明确需要证明的是作品复用、品类消费还是平台入口，再决定观察指标，避免用一个聊天角色的下降代表整个平台的衰退。
4. Prompt的持久价值，是把任务规则变成可维护的作品
线索：24:18–26:38、28:51–36:24。 创作者曾发现一种图片链接，只要在链接后放入画面描述，就能显示图片。他们教语言模型输出这种链接，让文本对话出现图片。另有人用一个总指令询问需求，再从五个工具指令中选择一个加载。
这些案例表明，语言可以作为组合能力的接口。但“五选一”的规则切换并不等于已经拥有自主规划与可靠执行的Agent。节目没有对这些玩法做稳定性测试。
Jay把Prompt类比为大模型的代码，强调提供场景、规则和信息，而非依赖神奇措辞。他认为随着模型能理解更多上下文，作品可承载更多信息；据其社区观察，模型变强时Prompt反而变长。这里的上下文是模型当次可处理的信息范围，长度增长并非质量提升的充分条件，节目也没有控制任务复杂度进行实验。
Prompt生成器因此可以降低实现门槛，像辅助写代码一样帮助人表达想法。Jay称当时尝试过专用生成模型，仍觉得通用模型配合指令更好；这是当期尝试结果。更长期的供给来自作者维护、二次创作和相互修改，而非一份永不更新的模板。他还提醒，投入最深的创作者未必有工程背景。
5. 商业化和融资，都要看完整过程
线索：37:22–38:08、38:51–48:57。 录制时FlowGPT还没开展商业化。Jay不赞成把Prompt本身作为买断商品，认可应用内增值付费。这是付费设计立场，尚未验证用户愿为哪些体验付费，也没有回答模型调用成本与分成怎样平衡。
融资故事同样容易被压成一句“几天拿到投资”。Jay讲到某基金在密集会面后迅速给出投资条款清单（TS），但对方此前已经研究市场与数据，他也已经与其他美元基金沟通近一个月。两天有效的条款还压缩了比较其他选择的时间。
编辑推论：可学习的是把研究、接触、决策和尽调拆开看。会面后快，不代表全流程短。关于中美机构、熟人引荐及数据库尽调的差异，只能保留为双方的经历与观察；不能据此概括全部基金，更不能作为融资或投资建议。
把本期变成一次创作者观察
以下是编辑设计的小实验，不是FlowGPT公开的运营方法。邀请少量目标创作者，各完成一个可以被别人直接使用的作品，连续观察两周：
- 记录从接触到首个作品发布，实际需要哪些人工帮助；尤其区分不会表达需求与平台能力不够。
- 请消费者完成一次具体体验，记录是否需要离站找零件、在哪一步中断。
- 复访时分别记录回到原作品、改看同类新作品、进入别的品类，并询问原因。
- 给作者一次修改机会，看用户反馈能否转成可见的版本改进。
如果只有首日尝鲜，就继续查体验与供给；如果用户换作品却持续回来，则进一步验证品类价值。小样本用来发现机制，不宜直接估算全平台留存。Jay结尾担心自己不再理解十六七岁用户的文化，也提醒团队：早期用户理解会过期，创始人年轻并不自动免疫。
回听导航与关联阅读
- 06:03：从社区逐个接触创作者。
- 12:27：GPT Store、开放生态与模型选择。
- 20:21：Prompt生成器、角色与文字冒险。
- 23:37：应用留存与品类留存。
- 29:33：Prompt会消失的部分与长期信息价值。
- 41:29：快速融资背后的提前研究。
相关主题：创作者供给、娱乐消费、低成本验证。可对照Hidecloud谈AI产品经理，比较任务表达与产品判断；再看EVE的关系设计，区分品类复访与关系连续性。
资料边界：依据完整提供的转写、旧精读及官方当期介绍重新组织，未逐句复听音频。姓名按官方校正；含混术语不直接引用。商业数字、竞品评价与预测保留说话者及时点，不当作当前已验证事实。
English edition
FlowGPT’s Early Growth: Let Creators Finish the Experience, Then Ask Why Users Return
42 Zhangjing episode analysis · AI creation platforms
Guest: Jay 党嘉成, introduced as FlowGPT’s founder | Host: Qu Kai
Published: March 14, 2024 | Duration: 49 minutes 57 seconds | Prepared: October 18, 2026
Recorded: November 2023; publication was delayed because of fundraising details
Official episode and introduction
Why a prompt community began calling itself an application platform
A prompt contains instructions, background and rules supplied to a model. Sharing one can resemble sharing a template, but a consumer may actually want a text adventure, a conversational character or a tool that repeatedly completes a task. FlowGPT’s changing description reflects a move from collecting instructions toward consuming those experiences directly on its platform.
The official introduction reports three million monthly active users as of November 2023 and announces a new funding round at the US$10-million scale around publication. These are statements by the program and participants, without independent audit, and do not describe today’s scale. The more useful questions concern organizing supply, understanding repeat visits and competing with model providers’ own application stores.
1. A two-day first version was followed by intensive personal contact
Episode reference: 03:13–08:37. Jay says the first version took two days. Seeing users spread it voluntarily on social platforms in the second week encouraged him to leave university and commit. Before launch, he had not identified one acute problem; he initially brought together people willing to spend time writing prompts.
This was more than an isolated inspiration. He had built projects since high school, read about products and entrepreneurship and accumulated collaborative experience at Berkeley. Early teammates included people whose reliability he had already observed. His claimed advantage of having little experience concerns having fewer commitments and a willingness to try quickly. It does not establish that experience harms entrepreneurship or that dropping out is a formula for success.
The labor came after release. He searched Discord, Reddit and Twitter for creators, reporting five or six hundred individual conversations in the first two weeks and thousands in total. Some people sharply criticized the product. Direct recruitment provided both supply and understanding: who would create seriously, where they got stuck and why they would share offered more specific information than a general question about liking AI.
Editorial interpretation: The transferable method is a low-cost prototype combined with concentrated user contact. Page views without published works and actual consumption do not establish that a two-sided platform has started functioning. This account also supplies no recruitment conversion rate or reproducible acquisition cost.
2. Model choice has to improve a creator’s finished experience
Episode reference: 13:00–16:26 and 19:35–22:56. Jay compares GPT Store to iOS and FlowGPT to Android. His intended strategy gives creators model choice and reduces dependence on one provider’s ecosystem. It rests on the assumption that consumers follow creators.
The analogy matters only when translated into an experience. Jay argues that model directories often send users elsewhere to deploy selected components. FlowGPT wants capabilities such as image generation embedded in a work, with consumption completed on the platform. A text adventure could produce scene images during the story, instead of requiring the player to find and deploy a model before returning. Bots are one content format within this vision.
The team tried making many applications itself, but concluded that committed creators understood prompts and play better. It therefore preferred distributing their work. Serving creators also includes exposure, responsive feedback channels and conditions for maintaining a work, beyond providing a model list.
Negative comments about Character.AI represent Jay’s account of creator interviews and the host’s response. They are not independently established findings about the competitor’s operation. Editorial interpretation: Whether openness becomes an advantage depends on model choice improving works, creators continuing to update them and consumers following. The episode provides insufficient data to establish a network effect.
3. Switching bots can still mean repeatedly consuming the same category
Episode reference: 17:04–18:38, 20:25–23:54 and 26:44–28:08. Jay distinguishes utility from entertainment. A utility user may repeatedly need the same function; an entertainment user might spend three days with one story before seeking another. Here a bot embodies an idea, much like a content item in a video service.
Examples include a prompt generator, companion characters and text games using generated images. Jay reports that around 60% of usage then related to entertainment. The interview does not specify the denominator, observation period or classification method. This cannot establish that entertainment dominates all AI use.
A more useful distinction concerns levels of retention. Some people return weekly to one application; others return weekly to find something new within a category. Changing works does not automatically mean the platform has lost value. Conversely, returning for novelty does not demonstrate a lasting relationship with a character.
Jay also reports people leaving the page open and returning whenever they want AI interaction. The host discusses competition for the user’s entry point. This is a directional signal without a measured proportion or duration. Editorial interpretation: Decide whether the claim concerns reuse of a work, category consumption or becoming a habitual destination before choosing metrics. Declining use of one character need not represent decline of the entire platform.
4. Durable prompt value turns task rules into maintainable works
Episode reference: 24:18–26:38 and 28:51–36:24. Creators discovered an image link that displayed an image when a description was appended. They instructed a language model to produce that link, bringing images into text conversations. Another creator built a coordinating prompt that asked what the user needed and loaded one of five tool prompts.
These examples show language functioning as an interface for combining capabilities. A five-way rule-based selection does not establish autonomous planning and reliable execution by an agent. The interview does not test these experiments’ stability.
Jay compares prompts with code for a large model, emphasizing scenarios, rules and information over magical phrasing. He expects models that understand larger contexts to support more information within a work. In his community observation, prompts became longer as models improved. Context means the information a model can process in the current interaction. Greater length is not sufficient evidence of better quality, and the episode offers no experiment controlling for task complexity.
A prompt generator can therefore lower the implementation barrier, helping people express ideas as coding assistance does. Jay says dedicated generation models they tried were then less effective than a general model with prompts; this remains a report of trials at that time. More durable supply comes from maintenance, derivative works and collaborative modification rather than a permanently static template. He also observes that the most committed creators need not have engineering backgrounds.
5. Commercialization and fundraising require the whole process to be visible
Episode reference: 37:22–38:08 and 38:51–48:57. FlowGPT had not begun commercialization when recorded. Jay opposes selling prompts themselves as a buyout product and favors additional purchases within applications. This is a pricing position, without demonstrated willingness to pay for particular experiences or a resolution of inference costs and revenue sharing.
The fundraising story is similarly easy to reduce to “investment arrived in a few days.” Jay describes a fund rapidly issuing a term sheet, or TS, following intensive meetings. However, it had already researched the market and data, while he had spent almost a month speaking with other dollar-denominated funds. A two-day expiry also restricted his opportunity to compare alternatives.
Editorial interpretation: Separate research, contact, decisions and due diligence. Speed after a meeting does not establish a short total process. Observations about Chinese and American institutions, warm introductions and database diligence reflect the participants’ experiences. They cannot characterize all funds or serve as fundraising or investment advice.
Turn the episode into a creator observation exercise
The following is an editorial experiment, not FlowGPT’s published operating method. Invite a small group of target creators to make one work that another person can use directly, and observe for two weeks:
- Record the personal assistance required between first contact and publication. Distinguish difficulty expressing a need from missing platform capabilities.
- Ask consumers to complete one concrete experience. Record whether they must leave to find components and where they stop.
- On return visits, distinguish returning to the same work, choosing a new work in the same category and moving to another category. Ask why.
- Give authors an opportunity to revise and observe whether user feedback becomes a visible improvement in the next version.
If activity consists only of first-day curiosity, investigate experience and supply further. If users switch works but keep returning, investigate category value. Small samples reveal mechanisms; they should not directly estimate platform-wide retention. Jay’s closing concern that he might stop understanding the culture of sixteen- or seventeen-year-old users is relevant here: early user understanding can expire, and a young founder is not automatically protected.
Listening guide and related reading
- 06:03: Recruiting creators through individual community conversations.
- 12:27: GPT Store, an open ecosystem and model choice.
- 20:21: Prompt generators, characters and text adventures.
- 23:37: Application retention versus category retention.
- 29:33: Which parts of prompting may disappear, and its continuing informational value.
- 41:29: Research preceding rapid fundraising decisions.
Related themes: creator supply, entertainment consumption and inexpensive validation. Compare Hidecloud on AI product management for task expression and product judgment, and EVE’s relationship design for the distinction between category return visits and relationship continuity.
Source boundary: Reorganized from the complete supplied transcript, earlier reading note and official contemporary introduction. Audio was not replayed sentence by sentence. The guest’s name follows the official introduction; unclear terminology is not directly quoted. Business figures, competitor opinions and predictions retain their speakers and historical context rather than becoming verified current facts.