Tencent, Alibaba, and ByteDance compete in the AI skills store market.

Skill is becoming one of the most popular keywords in the AI field.

Skill can be understood as the "operating manual" for an AI Agent. It is a structured instruction file that clearly specifies which tools to invoke, how to assess various situations, and the criteria for final output. The Agent reads this file and follows the predefined path to execute tasks.

For example, a seasoned product manager can encapsulate their entire process for writing product requirement documents into a Skill; any agent that installs it can generate a standardized requirement document using the same framework.

As the number of Skills increased, distribution platforms emerged. Initially, developer communities such as GitHub and ClawHub took on this role, with uploading, searching, and downloading of Skills all taking place within these technical communities.

Major companies are also rapidly catching up. In March this year, Tencent, Alibaba, and ByteDance successively launched Skill stores on their respective agent platforms. In the following two months, Zhipu, Meituan, and Xiaohongshu also entered the space. Internet giants, large model companies, local life industry leaders, and even content platforms are all vying for this entry point.

The essence of the Skill Store battle is positioning for the traffic gateway in the AI era—those who control distribution control the users.

But aside from ByteDance’s Coze testing paid skills, all other platforms are offering only free versions. If these “stores” aren’t profitable, why are companies still competing for them?

Who’s joining? Why is the Skill Store worth grabbing?

Before answering this question, first review a model that has already been successfully implemented.

In the mobile internet era, Apple’s App Store doesn’t just profit from its 30% commission on downloads; its core value lies in the fact that developers create apps to enter the iOS ecosystem, users remain within the iOS ecosystem to use these apps, and consequently continue consuming within the ecosystem—purchasing iCloud services, subscribing to Apple Music, and making in-app purchases. Distribution rights are the gateway; ecosystem consumption is the true source of revenue.

The Skill Store is competing for the same logic: users will remain within the ecosystem where they are accustomed to acquiring Skills and consuming services. The difference lies in the fact that this logic has already been validated in the mobile internet era, while the Skill Store is still in the "pie-in-the-sky" stage. Understanding this, we can now examine the different strategies of the three types of entrants.

The first category consists of major internet companies that drive traffic through the Skill Store and generate revenue within the ecosystem.

Ali has integrated the "Xiaoxiaobao" Skill marketplace into its JVS Claw Agent assistant, allowing users to sync selected Skills to their tools with a single click. The Skill marketplace itself is free, but using Skills consumes computing power, which generates revenue for Alibaba's cloud services.

ByteDance is pursuing two parallel paths. VolcEngine’s Find Skill targets enterprise customers, integrating Skills from multiple sources such as ClawHub and GitHub. Meanwhile, the Skill Store powered by Koubzi is designed for individual developers, lowering the barriers to creating and using Skills, and also enabling Skill monetization. The goal is to attract developers and drive cloud service and computing power consumption through Skills.

Tencent’s strategy is slightly different. SkillHub essentially serves as a localized mirror of ClawHub overseas, designed to drive traffic and adapt to local markets. However, Tencent’s true advantage lies in its WeChat Mini Program ecosystem. Leveraging mature service workflows built across millions of Mini Programs, Tencent can encapsulate various offline and online services into standardized Skills. If this path succeeds, the business model will resemble that of Mini Programs, generating revenue through transaction commissions and advertising.

Meituan, on the other hand, uses the Skill ecosystem to support its core business. In April, it launched xia345, positioned as an AI Agent ecosystem navigator, featuring over 20 Agents and more than 7,000 Skills. Shortly after, in May, it began public testing of Miyou, an AI community with over 3,000 registered Agents and more than 40,000 Skills total. Users discover content on “Miyou” and then download and use Skills via “xia345.” While Skills themselves do not generate direct revenue, they extend user engagement within Meituan’s ecosystem, creating more conversion opportunities for core businesses such as in-store services and food delivery.

The second category consists of large model companies that retain users through Skill stores and generate revenue through model calls.

In April, Zhipu launched the AgentMore Skills plaza on its own Agent platform, Auto Claw, integrating three modules—officially curated skills, Skill Hub, and open-source community—supporting one-click, zero-token installation.

The Moon's Dark Side acted earlier, launching Kimi Claw in February, allowing users to deploy Open Claw with a single click on the web platform and configure a skill library, enabling direct installation and invocation of various Skills within the browser.

It seems most natural for large model companies to distribute skills. The model itself serves as the foundation for running skills, and developing a skill store can drive continuous usage of their own large models, keeping users within their ecosystem.

He Yu, an Agent engineer at a large model company, mentioned that their self-developed Skills are better aligned with their underlying models, resulting in a superior user experience. Fundamentally, Skills are the "bait," while model invocation volume is the "fish."

The third category is content platforms that treat Skill as a new content category, earning revenue from traffic and advertising.

Xiaohongshu has recently launched Red Skill, which is currently in internal testing. Users can attach a Skill link beneath their posts, and clicking it copies the installation command directly. Unlike traditional Skill distribution, which follows a path from search to configuration, Xiaohongshu takes a content recommendation approach, transforming Skills into browseable, recommendable content formats. Xiaohongshu does not earn revenue from the Skills themselves, but rather from the traffic and advertising income generated by this content.

The logic is consistent across all three types of players: the Skill store itself does not generate profit, but it serves as the entry point to acquire and retain users. The real revenue comes from outside the Skill store.

However, this judgment holds true only if developers and users are genuinely willing to use it.

Independent developer blogger Sugimori Minami noted that the Skill stores embedded within big tech products may not be as attractive as imagined. They function more like secondary features within the overall product, with low visibility and not a primary focus for these companies. Meanwhile, content platforms inherently possess stronger dissemination capabilities, giving them a competitive edge in Skill distribution.

In other words, the store has been set up, but it still lacks sufficient appeal.

The most direct way to determine whether the Skill Store business is viable is to see if it makes money.

Currently, only ByteDance’s Koi supports Skill trading, allowing creators to set prices and sell their Skills. Almost all other platforms distribute Skills for free. The only real “trading” happening is individuals on Xianyu exploiting information asymmetry by bundling and reselling open-source Skills.

The "store" skill is still just a metaphor. What's the issue?

The first hurdle is that Skill is difficult to price.

The App Store succeeded thanks to a comprehensive evaluation system: clear functionality, stable experience, along with ratings and user reviews. More importantly, the same app delivers the same results for everyone.

What Skill lacks is this kind of certainty. Switching to a different model or context can lead to vastly different outcomes from Skill. Shan Sen nan told AIX Finance that performance varies across different Agent products, and the underlying model capabilities differ as well—resulting in unpredictable outputs for the same Skill across different products and models. Even within the same product and model, the inherent randomness of AI means outputs may still vary.

He Yu added another perspective: Most general skills designed for everyday users involve open-ended outputs without a single standard answer, and the industry currently lacks unified criteria for evaluating their effectiveness. High-quality skills cannot be effectively identified, resulting in extremely high screening costs for users.

If the results are inconsistent, an evaluation system cannot be established. Without an evaluation system, users lack a basis for payment.

The second hurdle is opaque costs.

The number of tokens consumed by different skills to complete the same task can vary by several times, but users have no way of knowing this before installation. There’s no way to compare which of two skills with the same functionality is more “token-efficient.”

Yu He gave an example: he once used two long-form Skills on the same platform to process the same document and issue identical instructions, yet the token consumption varied significantly—and this difference was entirely invisible when selecting a Skill. Users pay for Skills but still bear uncertain additional token costs—how should this be accounted for?

The third hurdle lies in security risks.

This year, there have been precedents of Skill poisoning attacks, where malicious Skills are uploaded by mimicking the names of popular Skills to steal user data. Although platforms have gradually implemented review mechanisms, this has also raised the barrier for developers submitting Skills.

When Sugimori Kusunoki tried uploading a Skill to Xiaohongshu, they encountered restrictions—the platform only allows uploading Markdown and TSD files, making it impossible to upload complex Skills in full, forcing them to downgrade it to a simple prompt. A balance between security review and developer experience has yet to be found.

The final hurdle is the lack of standardized protocols.

Different developers describe the same task in different ways, making it easy for the model to develop misunderstandings and resulting in inconsistent performance. He Yu noted that ambiguities in the descriptions make it difficult to control the actual user experience of Skill, turning “usability” into something mystical.

In addition, the lack of standardized permission boundaries prevents the ideal outcome of "develop once, distribute across multiple platforms" from being realized.

These four hurdles all point to the same underlying cause: Skill is inherently a personalized workflow that naturally resists standardization, yet commercialization requires standardization.

So, the current Skill Store is more like a display shelf—items are out there, but users don’t know which one to choose, and even if they do, they’re unsure if it will work well. There’s still a long way to go before it becomes a true “marketplace.”

First, shift your focus from the platform to the developers.

Independent developer Chen Xu previously uploaded a paid Skill on Kozhi. On the day it was approved, six people paid for it, and the homepage recommendation provided sustained exposure. However, this success was short-lived—he soon discovered that he no longer had a chance to appear on the homepage recommendation; users could only find his Skill by actively searching for it, and he was unable to run paid traffic campaigns. The opportunity for homepage exposure was entirely controlled by the platform and highly random.

This at least highlights two points: first, there is genuine demand for Skill payments; second, developers' distribution capabilities on existing platforms are extremely limited.

So, can the Skill Store become the next App Store? Currently, there are two obstacles.

On one hand, there is no standardized evaluation system for skills. Chen Xu mentioned that he typically selects skills based on GitHub star counts, as these reflect genuine user validation; however, popular rankings on domestic platforms differ from those overseas, leading to potentially distorted metrics. Without a cross-platform, standardized evaluation system, users are left to choose skills by chance.

On the other hand, Skills have strong personalized attributes.杉森楠 points out that most general-purpose Skills available on the market have limited effectiveness. Truly useful Skills must align closely with individual workflows, undergo repeated refinement in actual use, and eventually evolve into personalized methodologies. For example, even two Skills both labeled as “writing assistants” may differ entirely in their adapted workflows and output styles.

Without a proper evaluation system, the Skill Store can only remain at the stage of a display shelf.

But from another perspective, Skill is essentially a new form of product. In the past, users paid for “certainty”—if they needed a function, they downloaded an app. Now, they’re buying “possibility”—a capacity to create and a reusable methodology.

He Yu categorized scenarios with a payment foundation into two types: first, office essentials such as contract review and automated data report generation—fixed processes where enterprises have strong willingness to pay; second, personal tools such as resume optimization for job hunting and statement writing for study abroad applications—scenarios with relatively higher conversion rates for payments.

The question is, who can turn this space into a real business?

Each of the three types of participants has its own advantages, but also its own shortcomings.

Internet giants are closest to real-world scenarios, but Skill Stores are merely an add-on for them, and they won’t allocate core resources. Large model companies have a natural advantage in model adaptation, but their ecosystems can’t match those of internet giants; for them, Skill Stores are just a value-added service, with the core goal still being to encourage users to continuously invoke their models. Content platforms have the strongest dissemination capabilities—during the stage where Skill evaluation standards haven’t been established, users rely on influencer recommendations and usage demonstrations to choose Skills, which is precisely where content platforms excel. However, they are the furthest from the technical ecosystem.

The instability, personalized nature, and security risks of Skills make this business much harder than it appears on the surface. No one has yet made purchasing a Skill as natural as buying an app.

This article is from the WeChat public account "AIX Finance," authored by the AIX Finance team.