Obsidian Skills Teaches Claude to Master Knowledge Graphs

Obsidian Skills Teaches Claude to Master Knowledge Graphs

Kepano, a key Obsidian contributor, released obsidian-skills in January 2026 — a repo that teaches AI agents like Claude how to properly work with structured notes and knowledge graphs [4][5]. The skills follow the Agent Skills specification and enable agents to create Obsidian Markdown, use JSON Canvas, and navigate CLI operations [6].

Installation is straightforward: copy the skills to your .claude/ folder within an Obsidian vault. This bridges the gap between AI assistance and personal knowledge management, allowing agents to understand and manipulate your second-brain systems rather than just generating generic text responses.

Weaviate Tackles AI Agent Hallucinations with Specialized Skills

Weaviate launched agent skills in February 2026 to prevent coding agents from hallucinating when working with vector databases [7][8]. The open-source repo provides Claude, Cursor, and other agents with accurate, up-to-date instructions for Weaviate features including schema creation, hybrid search, and full application blueprints [9].

The initiative addresses a common problem: agents often generate outdated syntax or make up non-existent features when working with specialized tools. Agents can auto-discover these skills via npx install, ensuring they have reliable knowledge about vector database operations rather than guessing based on incomplete training data.

Karpathy Highlights Context Over-Reliance Problem in LLMs

Andrej Karpathy warned today about a fundamental bias in large language models: they over-rely on whatever context is provided, even when it's not the most relevant information [10]. As he noted, "LLMs develop a bias to use what is given [in context window during training], then at test time overfit to anything that happens to be in the context."

This observation has significant implications for RAG systems and memory features, where the retrieved context isn't always the most useful for answering a query. The post sparked discussions about improving retrieval-augmented generation and highlights why simply stuffing context windows with related information may not be optimal.

What This Means For Your Meetings

Today's developments reveal a clear trend: the infrastructure for intelligent knowledge work is rapidly maturing, but the challenge is shifting from raw capability to smart implementation. Insanely Fast Whisper democratizes high-quality transcription, while Obsidian Skills and Weaviate's agent capabilities show how AI can work more effectively with structured knowledge systems.

However, Karpathy's warning about context over-reliance is particularly relevant for meeting intelligence platforms. Simply retrieving and presenting related meeting snippets isn't enough — the system needs to understand which context is actually useful for the current query. This suggests that future meeting tools will need more sophisticated retrieval strategies, potentially ranking context by relevance rather than just similarity.

Key takeaway: The race is no longer about who can transcribe or search meetings fastest, but who can build the smartest context selection and knowledge synthesis capabilities.

Sources

1. https://github.com/Vaibhavs10/insanely-fast-whisper
2. https://modal.com/blog/choosing-whisper-variants
3. https://news.ycombinator.com/item?id=38266833
4. https://github.com/kepano/obsidian-skills
5. https://www.reddit.com/r/ObsidianMD/comments/1q8gn9c/kepano_released_obsidianskills_repo_what_custom
6. https://addozhang.medium.com/obsidian-skills-empowering-ai-agents-to-master-obsidian-knowledge-management-8b4f6d844b34
7. https://weaviate.io/blog/weaviate-agent-skills
8. https://github.com/weaviate/agent-skills
9. https://techrseries.com/artificial-intelligence/weaviate-launches-agent-skills-to-empower-ai-coding-agents
10. https://x.com/karpathy/status/2036841069636370467