I found a new time sink: building LLM wikis. I could roll my own, but for now I’m using a new experimental app called Wikiwise, which has been built by Tristan Homsi, one of Readwise’s founders. The app is free and open-source. (ETA: it’s also a native Mac app.)
From the Readwise changelog:
- NEW! Wikiwise (experimental) – Tristan built a native Mac app for setting up, customizing, and managing your own local-first markdown wiki with an AI agent. Just point it at a folder and your markdown files become interlinked pages. Add sources and the LLM agent reads them, writes summary pages, cross-references everything, and keeps it all consistent. The wiki compounds with every source you add, with no need for a database, configs, or account. Wikis can compile to a clean, fast website you can share on the web. If you decide to try it out, feel free to send us any feedback!
You can work with your wiki inside the app, but since it’s built on a folder of markdown files it creates in Finder, you can also open it up in an Obsidian vault.
The app Inspired by Andrej Karpathy’s llm-wiki pattern. The Antigravity Codes website has an illustrated explainer of Karpathy’s concept.
An aside: Karpathy’s LLM-wiki GitHub gist has already generated about a bazillion YouTube how-to videos if you want to roll your own. It’s very flavor-of-the-month … (There are also YouTubers recommending NotebookLM instead.)
You do need to be comfortable using Claude Code or its equivalent—there’s a terminal built right into the app—but nothing you need to do there is very complicated.
I’ve been plowing my way through a giant anthology of foundational essays on photography theory. I had Wikiwise ingest a markdown version of each essay in the anthology, and let it do its thing. The result was pretty interesting: Claude realized that it was ingesting an anthology rather than a random collection of articles and web clippings and proposed organizing the wiki around seven core concepts it found woven through the collection. Once I agreed, it summarized each of the core concepts; divided each concept into sub-categories; mapped each essay to one or more of those subcategories; and provided a narrative overview of each concept, its subcategories, how each essay mapped to that concept fit into the picture, noting material points of agreement/ disagreement.
One of the outputs is a set of recommended pathways through the collection, which I find particularly helpful.
And, you’re not locked into the app’s initial schema: you can work with Claude Code right inside the app to modify the way it works with the material you feed it or how it presents information. You can add new skills or modify the ones it has.
If you’ve got a folder full of markdown files on a particular topic and want to start building links between them, give it a spin. Start small if you’re just testing the concept, though, and save some tokens until you’re ready to commit to it.