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

SkillMarket presented an MVP of a domain skills marketplace for AI agents, built on FastAPI in just a few weeks. The project aims to solve an old problem with generative models: they're useful for general tasks, but in narrow professional scenarios without human expertise, they quickly start making mistakes.

Where the idea came from

The project team describes a scenario familiar to many companies: AI agents handle code, templated emails, and internal routine fairly well, but falter when they need to work in a domain with lots of nuances. The article provides examples from arbitrage and USN tax situations, where a "smart model" alone isn't enough. You need a sequential way of thinking like a practicing expert, otherwise the LLM starts hallucinating and producing dangerously inaccurate answers.

This gave rise to the hypothesis: instead of hiring an expensive senior specialist to sit alongside a developer and manually assembling chain-of-thought prompts each time, you can turn this expertise into a separate digital product. A lawyer, logistician, or other industry expert describes their working algorithm, and the business then purchases it as a ready-made skill for a corporate agent. Essentially, SkillMarket proposes moving prompt engineering from a development team craft into a separate market for domain expertise.

How the MVP is structured

The project's architecture is pragmatically designed without unnecessary layers. They chose Python and FastAPI for the backend, PostgreSQL and SQLAlchemy for storage, Celery and Redis for async tasks, and built the interface with Vanilla JS and Vite. The infrastructure is packaged in Docker Compose—the team deliberately didn't go for microservices for the MVP. This makes sense: the main value here isn't 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 calls to the LLM
• Vanilla JS, Vite, Sortable.js, and Marked.js — for lightweight frontend
• Docker Compose — for quick system deployment

The process of creating a skill was also made non-technical but "human." An expert doesn't need to know what few-shot prompting or JSON schema is: they simply enter the builder and freely describe in text how they solve the task in real work. After that, Celery sends this draft to Claude Sonnet via OpenRouter, and the model packages the thought flow into a structured system prompt with a name, description, purpose, tags, and token count estimate.

How quality is filtered

The next problem is moderation. If you simply open a marketplace to sell prompts, it will quickly turn into a warehouse of dubious templates with no real business value. That's why after publishing a skill, SkillMarket runs automatic checks right away through Claude Sonnet and GPT-4o. Both models assign scores across several criteria, and weak cards either go to manual review or get rejected outright.

• clarity — how clearly the instructions are described
• completeness — how fully the algorithm is covered
• domain_accuracy — whether there are obvious errors in domain logic
• reusability — whether the skill can be applied across different companies
• overall_score — summary score for publication decision

The team also thought through monetization separately. For payments, they integrated YooKassa: the buyer tops up their balance, 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 started reaching out. For an MVP, this is an important signal: the model is interesting not only to developers but also to practitioners themselves who hold applied expertise.

"Rent out your brains"

What this means

If a similar model takes off, the AI agent market could shift from selling "universal bots" to selling verified narrow-specialized skills. For business this is a convenient scenario: instead of growing your own prompt engineer inside each team, you buy ready-made expert blocks almost the same way you install libraries today through npm or pip.