Skill is becoming one of the hottest keywords in the AI field.
Skill can be understood as an "instruction manual" for an AI Agent. It is a structured instruction file that clearly states what tools to call, how to make judgments under specific circumstances, and the criteria for the final output. By reading this file, the Agent can execute tasks according to the preset path.
For example, a senior product manager can encapsulate their entire workflow for writing product requirement documents into a Skill. Anyone's Agent, upon installing it, can output a standardized requirement document following the same framework.
As the number of Skills increases, distribution platforms have emerged. The earliest players in this role were developer communities like GitHub and ClawHub, where Skills were uploaded, searched, and downloaded within the technical community.
Major tech companies are also catching up quickly. In March of this year, Tencent, Alibaba, and ByteDance successively launched Skill stores on their respective Agent platforms. In the following two months, Zhipu, Meituan, and Xiaohongshu joined the fray. Internet giants, large model companies, local service giants, and even content platforms are all vying for this entry point.
The essence of the battle for Skill stores is about securing the traffic gateway in the AI era. Whoever controls distribution holds the users.
However, except for ByteDance's Coze testing paid Skills, most platforms currently offer free versions. Why are companies competing for a "store" that doesn't make money?
01 Three Types of Players, Each with Their Own Agendas
Who's entering the fray? Why are Skill stores worth fighting over?
To answer this, let's look at a model that has already proven successful.
In the mobile internet era, Apple's App Store didn't just earn revenue from a 30% commission on downloads. Its core value lies in this: developers create apps to enter the iOS ecosystem, users stay in the iOS ecosystem to use these apps, and consequently continue to consume within the ecosystem: purchasing iCloud, subscribing to Apple Music, making in-app purchases. Distribution rights are the gateway, but ecosystem consumption is the revenue source.
Skill stores are competing based on the same logic. Where users habitually acquire Skills, they will stay and consume services within that corresponding ecosystem. The difference is that this logic has been validated in the mobile internet era, while Skill stores are still in the "pie-in-the-sky" stage. Understanding this, let's examine the different strategies of the three types of entrants.
The first type is the internet giants, using Skill stores to drive traffic and earn money within their ecosystems.
Alibaba has built the "Shrimp Little Treasure" Skill market into its JVS Claw Agent Assistant. Users can select Skills and sync them to the tool with one click. The Skill market itself is free, but using Skills consumes computing power, which translates to revenue for Alibaba's cloud business.
ByteDance is pursuing two paths simultaneously. Volcano Engine launched Find Skill, targeting enterprise clients, integrating Skills from multiple sources like ClawHub and GitHub. Coze's Skill Store targets general developers, lowering the barriers to creation and use, and even supports Skill sales. The goal is to capture the developer community, leveraging Skills to drive cloud service and computing power consumption.
Tencent's strategy is slightly different. SkillHub is essentially a localized mirror site of the overseas ClawHub, serving as a traffic driver and localization adapter. However, Tencent's real trump card is the WeChat Mini Program ecosystem. Leveraging the mature service links accumulated from millions of mini-programs, Tencent can encapsulate various offline and online services into standardized Skills. If this path succeeds, the business model would be similar to mini-programs, earning transaction commissions and advertising revenue.
Meituan is using the Skill ecosystem to bolster its main business. In April, it launched xia345, positioned as an AI Agent ecosystem navigation site, aggregating over 20 Agents and more than 7,000 Skills. Then in May, it publicly tested the AI community "Miyou," which hosts over 3,000 Agents and more than 40,000 Skills in total. From navigation to community, users discover shares on "Miyou" and go to "xia345" to download and use them. Skills themselves don't generate revenue, but they can extend user dwell time within Meituan's ecosystem, creating more conversion opportunities for core businesses like in-store services and food delivery.
The second type is large model companies, using Skill stores to retain users and earn revenue from model calls.
In April, Zhipu launched the AgentMore Skills Plaza on its Agent platform Auto Claw, integrating three modules: officially curated Skills, Skill Hub, and open-source communities, supporting one-click, zero-token installation.
Moonshot AI acted even earlier, launching Kimi Claw in February. Users can deploy Open Claw with one click on the web version and configure a skill library, allowing direct installation and invocation of various Skills within the browser.
For large model companies, distributing Skills seems most logical. The models themselves are the foundation for Skill operation. Developing Skill stores can drive continuous usage of their own large models, keeping users within their domain.
He Yu, an Agent engineer at a large model company, mentioned that self-developed Skills have higher compatibility with their underlying models, offering a better user experience. In essence, Skills are the "bait," and model call volume is the "fish."
The third type is content platforms, treating Skills as a new content category to earn traffic and advertising revenue.
Xiaohongshu recently launched Red Skill, currently in internal testing. Users can attach Skill links below posts, which can be copied for installation with a click. Unlike the traditional Skill distribution path of search-to-configuration, Xiaohongshu follows a content recommendation route, turning Skills into a content format that can be browsed and recommended. Xiaohongshu's earnings come not from Skills themselves, but from the traffic and advertising revenue generated by this type of content.
The logic for all three player types is consistent: The Skill store itself doesn't generate revenue, but it is the gateway to acquiring and retaining users. The real income lies beyond the Skills.
However, this premise holds only if developers and users are genuinely willing to use these stores.
Independent developer blogger Shan Sen Nan pointed out that these Skill stores embedded within major companies' products may not be as attractive as imagined. They are more like an accessory feature within the whole product, lacking strong presence and not being a primary focus for these giants. Content platforms, with their inherent dissemination capabilities, are more competitive in the Skill distribution phase.
In other words, the stores are built, but their appeal is still insufficient.
02 Where is the Business of Skill Stores Stuck?
The most direct way to judge whether the Skill store business is viable is to see if it makes money.
Currently, only ByteDance's Coze supports Skill transactions, where creators can set prices for their Skills. Other platforms are almost exclusively for free distribution. What could be considered a real "transaction" is actually happening on platforms like Xianyu, where people resell packaged open-source Skills, leveraging information asymmetry.
The "store" in Skill store is still just a metaphor. Where's the problem?
The first hurdle is that Skills are difficult to price.
The App Store succeeded because of a complete evaluation system: clear functionality, stable experience, along with ratings and user reviews. More importantly, the same App runs identically for anyone.
Skills lack this kind of determinism. Changing the model or the context environment can lead to vastly different Skill outputs. Shan Sen Nan told "AIX Finance" that the performance of different Agent products varies, and the capabilities of the underlying models differ. The results produced by the same Skill on different products and models are uncontrollable. Even within the same product and model, due to the inherent randomness of AI, the output may not be consistent.
He Yu added another perspective: most general-purpose Skills aimed at ordinary users involve open-ended output without a single correct answer. The industry currently lacks unified standards for evaluating effectiveness. High-quality Skills cannot be effectively identified, resulting in extremely high screening costs for users.
Unstable performance prevents the establishment of an evaluation system. Without an evaluation system, users lack a basis for payment.
The second hurdle is the lack of cost transparency.
Completing the same task, different Skills may consume several times more tokens, but users have no way of knowing before installation. For two Skills with similar functionality, which one is more "token-efficient"? There's no way to compare.
He Yu gave an example: he once used two long-text summarization Skills on the same platform. Processing the same document with the same instructions resulted in vastly different token consumption, and this difference was completely invisible when choosing the Skill. Users pay for a Skill and then bear the additional, uncertain cost of token consumption. How do you account for this?
The third hurdle is security risks.
There have been precedents of Skill poisoning incidents this year. Malicious Skills are uploaded by impersonating the names of popular Skills to steal user data. While platforms have gradually implemented review mechanisms, this also raises the threshold for developers uploading Skills.
Shan Sen Nan encountered restrictions when uploading a Skill on Xiaohongshu. The platform only allowed Markdown and TSD files, preventing the complete upload of complex Skills, which ultimately had to be downgraded to a Prompt. A balance between security reviews and developer experience hasn't been found yet.
The final hurdle is the lack of standardized protocols.
Different developers describe the same task in different ways, which can easily lead the model to misinterpretations, resulting in inconsistent execution performance. He Yu mentioned that ambiguity in descriptions makes the actual experience of Skills difficult to control, turning "ease of use" into a mystery.
Coupled with the lack of standardized permission boundaries, the ideal of "develop once, distribute across multiple platforms" has yet to be realized.
These four hurdles point to the same underlying reason: Skills are essentially personalized workflows, inherently resistant to standardization. And the prerequisite for commercialization is precisely standardization.
Therefore, current Skill stores are more like display shelves. The goods are out, but users don't know which one to choose, and after choosing, they don't know if it works well. There's still a long way to go before real "transactions" can happen.
03 How Far from Being the Next App Store?
Let's first shift focus from platforms to developers.
Independent developer Chen Xu once uploaded a paid Skill on Coze. On the day it passed review, six people paid, and homepage recommendation brought continuous exposure. But the good times didn't last. He soon found he no longer had the chance to appear on the homepage recommendation. Users had to actively search to find it, and there was no option to purchase traffic. The opportunity for homepage exposure was completely controlled by the platform, with strong randomness.
This indicates at least two things: First, there is real demand for paid Skills; Second, on existing platforms, developers have extremely limited distribution capabilities.
So, can Skill stores become the next App Store? Currently, there are two main obstacles.
On one hand, Skills lack a unified evaluation system. Chen Xu mentioned that he usually selects Skills based on GitHub stars, as these have been tested by real users. However, the popularity rankings on domestic platforms often deviate from those abroad, and metrics might be distorted. Without a cross-platform, standardized evaluation system, users can only choose based on luck.
On the other hand, Skills possess strong personalized attributes. Shan Sen Nan pointed out that the effectiveness of most general-purpose Skills on the market is limited. Truly useful Skills need to be closely aligned with personal workflows, repeatedly debugged in actual work, to distill a proprietary methodology. For example, even two "writing assistant" Skills might adapt to completely different workflows and produce distinct styles.
Without a functional evaluation system, Skill stores can only remain at the stage of being display shelves.
However, from another perspective, Skills are essentially a new form of commodity. In the past, users paid for "certainty"—they needed a function, so they downloaded an App. Now they are buying "possibility"—a creative capability, a set of reusable methodology.
He Yu categorized scenarios with payment potential into two types: First, office necessities, such as contract review, data report generation, and other scenarios with fixed processes, where enterprises have strong willingness to pay. Second, personal tools, such as job resume optimization, study abroad essay writing, where the payment conversion rate is relatively higher.
The question is, who can turn this potential into a real business?
The three types of entrants each have their advantages and shortcomings.
Internet giants are closest to real-world scenarios, but Skill stores are merely "add-ons" for them, unlikely to receive core resource investment. Large model companies have a natural advantage in model compatibility, but their ecosystems can't match the giants. Skill stores are just value-added services; the essence is still hoping users continuously call their models. Content platforms have the strongest dissemination power. In the current stage where Skills lack a standardized evaluation system, users rely on blogger recommendations and usage demonstrations to choose Skills—something content platforms excel at. However, they are the farthest from the technical ecosystem.
The instability, personalization, and security risks of Skills make this business much harder than it appears on the surface. For entrants, making "buying a Skill" as natural as "buying an App" is a goal no one has achieved yet.
This article is from the WeChat public account "AIX Finance", author: AIX Finance Team
