Loading today’s stories
Product1 distinct publisher3 min readUpdated
Jinguyuan's owner made his menu and wait times readable by other people's AI agents. He says it draws more press than customers, which is the interesting part.
The Product Desk · Product desk
Compiled by The Product DeskSomething wrong?How this is made
Jinguyuan, a dumpling shop in Beijing, has published an agent "skill" that lets a customer's personal AI assistant pull up the menu, ask for a recommendation, or join the queue [1]. It is the cheapest example so far of a storefront treating agent-readability as a distribution channel, and the person asking that question is a 41-year-old restaurant owner rather than a platform [2].
Worth being precise about what was built. A skill, in this context, is a file that teaches an AI agent a specific workflow [3]. There is no app, no integration contract, no marketplace listing to negotiate. The owner, Li Bo, made it himself in April after watching agentic tools spread in China [2]. He also vibe-coded a website that tracks queue status at Jinguyuan's two branches, mostly with Alibaba's coding tool Qoder [4]. In July, roughly three months later, he partnered with an AI cloud provider to give every customer a token coupon worth 10 yuan, about $1.50, that could go toward model subscriptions [5][6].
Li told Rest of World the skill brings more media attention than actual usage [7]. That admission is the load-bearing fact here, not a caveat to skip past. The channel does not work yet. The cost of occupying it was one file and some weekend coding, which is exactly the asymmetry that produced a decade of speculative search-engine optimisation: when the price of being indexed is near zero and the payoff is unknown but potentially large, somebody builds early and badly. Li's own framing is the same argument in older clothes. "Someone needs to start using it, even when it's not that good," he said, comparing early steam engines being mocked as worse than horses to the eventual arrival of the car [8]. He also quoted Deng Xiaoping's line that science and technology are a primary productive force [9].
The demand side is not imaginary. A 2025 Ipsos survey found that of 30 countries polled, the Chinese population reported the highest level of excitement about AI [10]. When the agentic tool OpenClaw went viral, people queued to have it installed on their laptops [11]. Rest of World's reporter contrasts this with New York, where small businesses rarely advertise how much AI they use even when their menus and posters show the signs [12]. Li's route to this is not a tech founder's: he took a master's in telecommunications engineering in 2010, then worked at hostels and sold homemade desserts before opening Jinguyuan with a dumpling chef near his alma mater, serving students and tech workers [13][14].
Three things to watch. First, whether agents actually resolve these files without a directory, because a skill nobody can discover is a robots.txt entry for a site with no inbound links. Second, who ends up owning the index, since maps, delivery apps and model providers all have a claim on being the layer that decides which dumpling shop an agent recommends. Third, maintenance: a menu file that goes stale is worse than no file, and the operator cost of keeping it accurate is the thing no one has priced. Li's queue tracker is the tell, since live wait times are the part an agent cannot guess and a human must keep true [4].
Follow any of these and your For You feed starts watching them — no settings page required.
Ranked by verification strength, evidence, and original report placement.
Jinguyuan, a dumpling shop in Beijing, created a "skill" that allows users to check the menu, seek recommendations, or join the queue by talking to their personal AI agents.
The shop owner, 41-year-old Li Bo, said he built the AI skill in April as he witnessed the growing popularity of agentic AI tools.
Li vibe-coded a website tracking the queue status at Jinguyuan's two branches, mostly with Alibaba's coding tool Qoder.
In July, Li partnered with an AI cloud provider to give every customer a token coupon worth 10 yuan ($1.50) that could be put toward model subscriptions.
Li admits his AI skill brings more media attention than actual usage, though he is proud of becoming a role model for AI adoption.
Evidence-backed comparisons of source perspectives and observed adoption signals. Read the methodology
Which Builder, Operator, and Investor concerns the observed source mix emphasized—not a truth score.
Evidence, demonstrated adoption, hype gap, incentives, and confidence are assessed independently, each on its own current evidence. How these are measured.
Single-source first-person reporting with named owner, no independent verification
All material comes from one Rest of World dispatch built on a direct interview with the named shop owner, plus one externally citable data point (the 2025 Ipsos 30-country survey). The build details, dates, and coupon partnership are specific but rest solely on the owner's account; the cloud partner is unnamed, dates are month-level, and no usage metrics, code, or skill artifact is shown. That supports the existence of the experiment but not its scale or effect.
One shop, two branches, owner-acknowledged low actual usage
Concrete deployment exists -- a shipped skill file, a live queue-status site across two branches, and a July token-coupon promotion -- but the operator himself says the skill draws more press than use, and no order, session, or customer counts are disclosed. Broader signals (OpenClaw install queues, Ipsos excitement) show ambient consumer interest in agents rather than adoption of agent-readable storefronts.
Slightly overstated at the ecosystem level, but the source discloses the gap itself
The 'poster child of AI adoption' framing and the national-enthusiasm data invite a larger inference than one under-used skill file at a two-branch dumpling shop can carry. The gap is small rather than large because the reporting explicitly surfaces the shortfall -- the owner concedes press exceeds usage -- and adds a deflating contrast with New York shops that use AI without advertising it. Residual overstatement comes from generalizing a single anecdote into a national adoption pattern.
Owner gains publicity and a vendor partnership; tooling vendors gain a showcase
The subject has a clear stake: the skill demonstrably produces media coverage for his restaurant, he says he is proud of being an adoption role model, and he entered a promotional partnership with an AI cloud provider that distributes model-subscription credits to his customers. Alibaba's Qoder benefits from being named as the tool behind the build. These incentives are visible in the source rather than inferred, but the source does not disclose payment, sponsorship, or the partner's identity, so the strength of the commercial pull cannot be sized.
Directionally credible pattern, weak quantitative and corroborative footing
Confidence is moderate-low: one publisher, one interviewee, month-level dates, an unnamed commercial partner, and no usage or financial figures. What is reliable is that an agent-readable storefront surface was built and deployed and that the operator judges it under-used; what is not reliable is any claim about scale, effectiveness, or how representative it is of Chinese small-business AI adoption.
build
China's accelerator swap makes Cambricon supply, not export policy, your ship-date risk1 distinct publisher
build
AI-written code fails the same four ways, and every gate you own reports green1 distinct publisher
invest
Unitree's $905M Shanghai listing prices humanoids at 35x sales while profit halves1 distinct publisher
invest
The chips never move: Washington's fix for the Southeast Asia compute loophole1 distinct publisher
Distinct publishers with included, body-backed reporting in this cluster.
1 article · August 21, 2026