This combo makes smart note systems magical.
What this covers
Van Eyck walks through how AI can amplify a personal knowledge management system, specifically Obsidian, though the scope extends beyond coding. The video is structured around three working demonstrations: semantic search across a vault, interactive quizzes generated from notes, and batch-processing slide decks into markdown. The underlying claim is that while AI coding assistants deliver only 10–30% gains on real large-codebases (not the advertised 10x), the genuine 10x productivity wins happen in non-coding knowledge work—particularly in the tedious manual tasks that plague anyone managing a second brain.
The first demo shows semantic search: instead of matching keywords, the system searches by meaning using embeddings stored in a vector database, letting you find notes conceptually related to a query. The second builds an on-the-fly quizmaster that pulls from your vault, generates questions tailored to your notes, scores answers, and adds unrequested features like hints—all driven by an LLM agent working within a Zettelkasten structure. The third tackles the slide-deck problem: feeding hundreds of exported PNGs to a multimodal LLM to OCR and summarize in roughly 20 minutes, a task that would consume four hours by hand. Throughout, Van Eyck emphasizes a critical boundary: the LLM augments your thinking but does not do it. New content added to the vault is constrained to text directly from the slides and shaped by system prompts derived from analyzing the speaker's own writing patterns. Implementation details are included—TypeScript, Claude with an MCP server, Chroma—and the code is available on GitHub.
The speaker argues that the much-hyped '10x' gains from LLM coding assistants are overstated for actual coding (closer to 10-30%), but that genuine 10x productivity gains do exist in non-coding knowledge work like semantic search, learning via quizzes, and reverse-engineering legacy content within a personal knowledge management system.
- Real coding gains on large codebases are only 10-30%, not 10x
- LLM agents act as a 'programmable machine' you program in English once you accept their quirks
- The biggest 10x gains come from tedious non-coding tasks like converting 300-slide decks into markdown summaries
This asset isn't compiled yet
You're seeing its claims, ranked. Compile it to build the argument threads, weight them, and check each claim against your library — the full view.
The most effective way to learn is spaced and interleaved retrieval practice — mixing up topics with question/answer cards where well-known items move to the back and struggled-with items stay at the front — a principle drawn from the book Make It Stick.
“the theory is the best way to learn is to do spaced and interled retrieval practice which is a very fancy term of just mixing up the topics and having like little cards with questions and answers and like the the ones you know well go to the back of the stack. The ones you struggle with stay at the front.”
An LLM agent can generate interactive on-the-fly quizzes from your own notes by running a semantic search over the vault, building questions, scoring answers, and — because it is an LLM — improvising features like hints that were never explicitly programmed.
“I did not program anything about a hint, but seeing that this is an LLM, it can just like interpret my command and start freewheeling.”
LLM agents are effectively a new kind of programmable machine that you program in English, requiring you to deal with quirks like non-determinism and imperfect instruction following before you can get useful work out of them.
“these LLM agents are kind of a new programmable machine where you like program in English and you have to deal with all kinds of quirks like the non-determinism, the meh so so um instruction following stuff like that. But once you like get used to those quirks, you can actually get them to do useful stuff.”
A semantic search system can be built for Obsidian by vibe-coding a TypeScript plugin that calls Claude Code, which uses an MCP server to query a Chroma vector database storing embeddings of every note, enabling similarity search on meaning rather than keywords.
“I vipcoded a plugin that would u look at the file I have open and do like a semantic search... this plug-in just calls cloud code... It has a vector database in the background... I used Chroma DB... I installed an MCP server in my cloud that allows the agent to talk to this uh vector database.”
Embedding models work by converting a document into a vector of numbers that represents the document's meaning, a mechanism similar to the internal embeddings used inside large language models, which then enables similarity search across stored documents.
“given a document like one of my uh obsidian nodes just converts that document into a number sequence of numbers actually a vector uh which is that's like called an embedding. It's a vector representing the meaning of the document. It's it's quite similar actually to the internal mechanisms of a large language model.”
Using multimodal LLMs (e.g., Claude Code with image input) to batch-OCR and summarize hundreds of PowerPoint slides exported as PNGs into a single markdown note delivers genuine 10x time savings — reducing roughly 4 hours of manual work to about 20 minutes of cross-checking.
“it's like 300 slide 300 slide deck condensed into a single markdown file and I just have to scroll over both once to do a cross check which is like 20 minutes work instead of 4 hours of work. So this is one of those places where I really get 10x improvements u during my non-coding activities.”
When letting an LLM agent into a personal knowledge/second-brain system, you must not let it do your thinking or generate new content, because the system's value comes from augmenting your own brain — letting an LLM 'word vomit' into it destroys that value.
“when you let uh an LLM agent into your vault, into your smart node system, do not let it think for you. This second brain thing is like made to augment your brain. But if you just let an LLM uh vomit word vomit all over it uh without you having any input, you lose all all value of this uh system.”
AI coding assistants do deliver genuine 10x gains in rapid prototyping — going from idea to a hacked-together web app to showcase or explore alternatives.
“rapidly prototyping stuff like from idea to a hack together web app to showcase something or to explore alternatives. Sure, the 10x gain is there.”
The widely advertised 10x improvement gains from AI coding assistants are exaggerated; in actual coding on large codebases the gain is more like 10% to 30% on a good day, not 10x.
“the gains are there, but they are nowhere near 10x, especially during coding. Uh when I work on large code bases, it's like more like a on a good day 10% to 30% gain.”
To preserve fidelity and personal voice when the LLM writes the only content it is allowed to add to the vault, the speaker constrains it to text found only in the slides and supplies a system prompt derived from having Claude analyze his entire vault for writing patterns and tone of voice.
“I made like 200% sure that it writes only text that's found in the slides and that it like writes in my style. So, I had uh Clal analyze my entire vault for like writing patterns, tone of voice... and I had it come up with a system prompt like with the the summary of my writing style and that I also pipe in when I have it write stuff.”
The Zettelkasten / smart-notes system functions as a second brain: an offline, hyperlinked copy of one's thinking where notes reference other notes, strengthening thinking, writing, and the creative process.
“it's my second brain. It's like an offline copy of my brain... It's also hyperl so notes reference other notes and you can traverse the graph like that. uh and that's a I found it to be a really powerful way to uh strengthen my thinking and my writing and my creative process.”
LLM-generated drafts of courseware are not perfect and should be proofread before being pushed to production, even when they correctly parse formulas, quotes, and dense text from slides.
“This is something I would very much proofread uh before I would push this to production”
Obsidian's built-in search is limited because it is purely text/keyword-based (file name, tag, keywords), so searching for notes 'about LLMs' only matches the literal text and cannot find semantically related notes.
“it gives you like search on a file name on a tag on keywords uh stuff like that but it's all like text based heavy based stuff... it's pretty like just looking for the text LLMs and it's not a smart search or a semantic search.”
The entire semantic search system took roughly half a day to build (about two hours of setup plus an hour of system-prompt tweaking) and runs at no cost, since it is a self-built free plugin.
“I think it took me two hours to set up everything and maybe an hour to tweak the the system prompt... So this was half day's work and now I have semantic search in Obsidian.”
Running Claude Code with 'dangerously skip permissions' should only be done if you know what you are doing.
“Always run in and dangerously skip permissions if you know what you're doing. And if you don't, please don't touch this.”