Everything Is Becoming a Skill: AI Is Now the One Using Your Service

According to Mintlify, the developer docs hosting platform, AI now accounts for nearly 50% of all documentation reads, up from just 15% a year ago. Supabase’s numbers are starker: last week, bot visits to their guides outnumbered human visits three to one.

The shift is real, but what it means for service businesses is still being worked out. Downloads can explode while revenue collapses. That tension is where things get interesting.

What Happened to Tailwind

Tailwind CSS, widely regarded as the gold standard of CSS frameworks, announced a 75% workforce reduction two weeks ago. AI coding agents drove monthly downloads to 75 million, but almost no one visited the official docs anymore. Traffic dropped 40% and revenue dropped 80%.

Downloads exploded while revenue collapsed. The traditional model, where documentation traffic and developer mindshare translate into commercial adoption, broke down when the readers became bots. Tailwind’s case is the clearest example of this failure mode, but it is not the only one. Any service whose business depends on human attention to documentation is exposed to the same dynamic.

Anthropic’s Opening Move

The turning point came when Anthropic released an official frontend design skill and opened a plugin ecosystem where skills could be shared freely. Fully automated agent systems like the Ralph loop got packaged as skills. Adoption spread rapidly beyond development into non-technical domains. The organic growth pattern that once defined n8n templates started repeating itself.

The difference this time is that skills route AI directly to specific services, bypassing the open web entirely.

How Services Are Responding

After seeing what happened to Tailwind, services started packaging their core knowledge as agent skills. Vercel released vercel-react-best-practices, a comprehensive React knowledge base that naturally integrates their deployment and logging pipeline. Supabase published postgres-best-practices covering PostgreSQL performance, security, and schema design. Stripe, RevenueCat, Expo, Neon, and Better Auth all followed with their own skills covering payment integration, in-app purchases, React Native development, serverless database operations, and authentication setup.

Vercel’s approach is worth examining closely. Their skill makes it possible to build without deep React knowledge, and in doing so naturally funnels AI toward their deployment ecosystem. The skill creates value for the agent while quietly establishing Vercel as the default infrastructure choice. Whether that conversion holds long-term, whether agents start recommending alternatives once they have enough context, is an open question.

Beyond Text: Video Enters the Picture

Remotion released a skill recently, opening up territory that had not been touched yet. Known among developers as the framework that makes TikTok-style video editing possible through code, Remotion’s move signals that skills are not limited to text and code generation. Video editing, content production, and design are now in scope.

Building My Own

I built two skills recently: an agent-browser skill and an Excalidraw-based diagramming skill. Turning diagrams into images had always been a bottleneck in my workflow. The Excalidraw skill let AI visit Excalidraw on my behalf, do the work, and return the result. What started as a personal inconvenience clarified something about how skills actually work at the individual level: they are not just distribution tools for companies. They are automation handles for anyone who has a repeating workflow with a tool they do not fully control.

Who Still Decides

People are no longer searching and reading documentation directly. But context still requires human knowledge, and the decision to purchase a service still belongs to humans. The skill that an AI uses does not determine what a person buys. It shapes the conversation that leads to that decision.

That gap between AI consumption and human purchasing is where the real strategic question sits. The services that build skills now are betting that influencing AI’s defaults today translates into commercial preference later. That bet might be right. It might also be more complicated than it looks once AI agents start reasoning about vendor trade-offs rather than just following embedded recommendations.

What is clear: if AI is already your primary reader, waiting to build a skill means ceding that channel to competitors who have already shipped one.