SkillMarket builds a marketplace of skills for AI agents on FastAPI and Claude Sonnet

SkillMarket has shown an MVP of a marketplace for domain-specific skills for AI agents, built on FastAPI in just a few weeks. The project tackles an old problem with generative models: they are useful for general tasks, but in narrow professional scenarios they quickly start making mistakes without human expertise.

Where the idea came from

The team describes a scenario familiar to many companies: AI agents handle code, templated emails and internal routine reasonably well, but fall apart when they must work in a subject area full of nuance. The article gives examples from arbitration and from taxes under the USN (the Russian simplified tax system), where a "smart model" alone is not enough. What is needed is the consistent way of thinking of a practicing expert; otherwise the LLM starts to hallucinate and produce dangerously inaccurate answers.

Hence the hypothesis: instead of seating an expensive senior specialist next to a developer every time and assembling chain-of-thought prompts by hand, that experience can be turned into a separate digital product. A lawyer, logistics specialist or other industry expert describes their working algorithm, and a business then buys it as a ready-made skill for a corporate agent. In essence, SkillMarket proposes moving prompt engineering out of the development team's craft and into a separate market for domain expertise.

How the MVP is built

The project's architecture is deeply pragmatic, with no unnecessary layers. The team chose Python and FastAPI for the backend, PostgreSQL and SQLAlchemy for storage, Celery and Redis for asynchronous tasks, and built the interface in Vanilla JS with Vite. The infrastructure is packaged in Docker Compose, so the team deliberately did not go with microservices for the MVP. That looks logical: the main value here lies not in a distributed platform but in the speed of testing the hypothesis itself.

• FastAPI and OpenAPI: for the API and future integrations
• PostgreSQL, SQLAlchemy and Alembic: for data and migrations
• Celery and Redis: for background LLM calls
• Vanilla JS, Vite, Sortable.js and Marked.js: for a lightweight frontend
• Docker Compose: for quickly deploying the whole system

The skill-creation process is also designed to be "human" rather than technical. An expert does not need to know what few-shot prompting or a JSON schema is: they simply open the builder and describe in free text how they solve the task in real work. Celery then sends this draft to Claude Sonnet through OpenRouter, and the model packages the stream of thought into a structured system prompt with a title, description, purpose, tags and an estimate of token volume.

How quality is filtered

The next problem is moderation. If the storefront is simply opened for selling prompts, it will quickly turn into a warehouse of dubious templates with no real value for business. So after a skill is published, SkillMarket automatically runs a check through both Claude Sonnet and GPT-4o. Both models score it on several criteria, and weak listings are either sent to manual review or rejected at the outset.

• clarity: how clearly the instructions are written
• completeness: how fully the algorithm is covered
• domain_accuracy: whether the domain logic contains obvious errors
• reusability: whether the skill can be applied across different companies
• overall_score: the aggregate score used to decide on publication

The team has also worked out monetization. Payments run through YooKassa: the buyer tops up a balance and purchases a skill, the platform keeps a 20% commission, and the rest goes to the author's internal account. The article's author writes that after the first publication about the project, not only IT specialists but also experts from other fields began to show up. For an MVP this is an important signal: the model appeals not only to developers but also to the practitioners who hold applied expertise.

"Renting out your brains"

What this means

If such a model takes off, the AI agent market could shift from selling "universal bots" to selling proven, narrowly specialized skills. For business this is a convenient scenario: instead of growing a prompt engineer inside every team, companies buy ready-made expert blocks, much as they install libraries through npm or pip today.