LLM-Powered Personal Knowledge Bases with Obsidian

Obsidian's reach is substantial: the platform counts over 1.5 million users with 22% year-over-year growth, more than 2,700 community plugins (of which 100+ are AI-focused), and an average daily usage of 43 minutes per user. These figures paint a picture of a deeply engaged user base that spends meaningful time inside the tool — making it a high-value target for AI agent integration.

The social signal around this trend has been extraordinary. Karpathy's X post describing his AI-maintained wiki workflow received 29K likes, making it one of the most viral AI productivity posts in recent memory. Elvis Saravia (@omarsar0) shared his own approach to paper curation and Obsidian vaults with interactive artifact generators and MCP tools, earning 2.9K likes. Steph Ango noted via X that 'More and more people are using Obsidian as a local wiki to read things your agents are researching and writing,' garnering 1.7K likes. On YouTube, Greg Isenberg's video 'How I Use Obsidian + Claude Code to Run My Life' hit 242K views, Nick Milo's 'Obsidian + AI: How to Do It The Right Way' introducing the IDI Framework reached 115K views, and Zen van Riel's tutorial on letting Claude automate Obsidian notes pulled 56K views. This cross-platform attention signals that the topic has broken out of developer circles into the broader productivity and knowledge-work audience.

The release of Obsidian Skills marks an architectural departure from how software tools have traditionally integrated with AI. Rather than embedding AI as a feature within the app (the 'Ask AI' button pattern adopted by most SaaS tools), Skills provide AI agents with portable instruction sets that describe how to interact with Obsidian vaults. As one analysis explains, 'Skills are not plugins. They don't require migrating your data into a proprietary app database. They're portable rulebooks, really.'

This approach has several structural advantages. First, skills are model-agnostic — they work with Claude, GPT, or any agent that can read markdown instructions. Second, they preserve Obsidian's local-first philosophy: data stays on the user's machine as plain files while the AI agent operates on it through defined interfaces. Third, skills are composable and open-source, meaning the community can create domain-specific skill sets for different workflows. Addo Zhang argues this will extend well beyond note-taking: 'vertical domain skills will become the standard mode for deep integration between AI agents and professional tools.' If correct, Obsidian Skills may serve as a template for how the entire software industry bridges the gap between existing tools and AI agents.

Karpathy's vision is ambitious. In his viral X post, he described a workflow where 'the LLM writes and maintains all of the data of the wiki and he doesn't manually edit/add anything,' and argued that 'there is room here for an incredible new product instead of a hacky collection of scripts.' This framing — AI not as assistant but as primary knowledge worker — resonated because it articulated what many power users were already attempting in piecemeal fashion. Eric Ma's practical results validate the productivity gains: he reports knowledge management overhead dropping from 30-40% to under 10% of his time. But he also provides a reality check on reliability, noting that 'hallucinations occur maybe once every four or five sweeps, and usually trace back to inaccurate transcripts.'

The risks are real and documented. The Stack Overflow blog warns that AI-driven second brains come at a cognitive cost, reporting roughly 76,000 problematic AI interactions daily at scale. Security researchers have documented common vulnerabilities in AI agent architectures, highlighting that local file access by AI agents introduces attack surfaces that traditional note-taking apps never had to consider. On the enterprise side, Gartner's prediction that 40% of enterprise applications will embed task-specific AI agents by end of 2026 suggests the pattern Obsidian Skills pioneered will face intense scaling pressure. The question is whether the open, local-first, skill-based approach can compete with proprietary alternatives that offer tighter integration but less user control.