Everyone was adding AI to their product in 2024. The pattern was familiar: slap a chat interface on the homepage, call it an “AI assistant,” ship it before Q2 planning. We decided to ask a different question first.
Not “how do we add AI?” but “what would it mean for AI to actually understand Skillshare?”
That distinction sounds philosophical. It had very concrete consequences.
The Search Problem Nobody Wanted to Admit
Here’s what Skillshare’s search looked like before: type “watercolor techniques,” get classes about watercolor techniques. Type “I want to paint flowers like Monet” — get nothing useful. Maybe a class about Impressionism. Maybe nothing at all.
This isn’t unique to Skillshare. It’s how most search works. It’s matching words to words, not meaning to meaning.
The old system was like a library card catalog from 1987. It was organized, it worked for exactly what it was designed for, and it completely fell apart the moment someone walked up and said “I’m looking for something that made me feel the way I felt reading The Alchemist.” The catalog doesn’t understand feelings. It understands subject headings.
A good librarian does something different. You describe what you’re looking for — vaguely, conversationally, maybe with a metaphor — and they translate that into something the shelves can answer. They understand what you mean, not just what you said.
That’s the gap we were building toward closing.
What We Built: The Search Architecture
We designed a hybrid retrieval architecture for Skillshare’s class discovery:
Vector search converts text into mathematical representations of meaning. When you type “I want to paint flowers like Monet,” the system understands that this relates to watercolor, Impressionism, color theory, and botanical illustration — even if none of those words appeared in your query. It finds classes that are semantically close to your intent.
BM25 lexical search handles the cases where exact terms matter — instructor names, specific technique names, proprietary terminology. Sometimes you do want the exact keyword match.
A reasoning layer (RAG pattern) sits on top: a language model that reads the retrieved results and synthesizes a response grounded in actual Skillshare content. Not hallucinated summaries. Real classes, real teachers, real projects — surfaced in a way that answers the actual question.
🚧 Need more context: Which LLM model powered the reasoning layer? What was the production rollout timeline — phased by feature or by user segment? Any early adoption or search quality metrics we can include here?
The MCP Layer: Making AI Actually Know Where It Is
Here’s the part that most AI implementations skip, and it’s why so many of them feel hollow.
An AI that can search the internet but doesn’t know it’s operating inside a creative learning platform isn’t that useful. It’ll give generic answers. It won’t know the difference between a beginner illustration class and an advanced typography deep-dive. It won’t understand that Skillshare’s classes are taught by working creatives, not academics.
Model Context Protocol (MCP) is the solution to that problem. Think of it as the briefing document AI systems read before they do anything.
We designed an MCP layer that gave AI systems structured, safe access to Skillshare’s actual data: the class catalog, creator profiles, skill taxonomy, learner project history, category relationships. When an AI system operates inside Skillshare with MCP, it knows what platform it’s on, what content exists, and what the relationships between things mean.
This matters for two things. First, it improves quality — an AI that understands the platform gives better, more grounded answers. Second, it enables safe external integrations — if an outside AI agent wants to query Skillshare data, MCP controls what it can see and how it can use it.
The Principle That Held Everything Together
Skillshare’s creative community has complicated feelings about AI. For good reason: a lot of AI tools in 2024 were positioned as replacements for the kind of human creative work Skillshare’s teachers make a living from.
We held a clear line throughout this work: every AI recommendation surfaces a real human teacher. The AI helps learners navigate. It doesn’t generate the class. It doesn’t replace the creative. It’s the librarian, not the author.
That design principle wasn’t just ethical positioning — it was right for the product. The value of Skillshare is the human expertise and creativity inside it. AI’s job was to make that more accessible, not to substitute for it.
What Changed
The architecture established a foundation for natural-language discovery that goes far beyond keyword matching. Long-tail classes — niche topics, emerging skills, subjects that learners couldn’t find before because they didn’t know the exact right words — became genuinely discoverable.
🚧 Need more context: Any specific metrics on search quality improvement, long-tail class discovery rates, or user satisfaction scores from the semantic search rollout?
What I’m most proud of isn’t a single feature. It’s that we built something that could keep getting better — a context layer and retrieval architecture that compounds in value as the platform grows.