Last weekend 25th to 27th of September, members of Wikimedia Slovensko met in Dubnica nad Váhom. It was great to meet the fellow Wikipedians in the flesh again (Matej, Patrik and Filip), and to meet Kormi in 3D for the first time.
It was mainly a planning, coordination and brainstorming meeting, but part of the programme happened outdoors, with cameras. By Sunday the town’s 18th-century Calvary had its own photos on Wikimedia Commons, a record on Wikidata and the start of a Slovak Wikipedia article. The local Jewish cemetery gained a set of new photographs too. That weekend is a good small example of what this whole thing is about.
Plain markdown files turned out to be the best possible interface between a decade of notes and a language model. No API, no export, no connector, no sync service in the middle: the agent opens the folder and reads it, the same way I do. This post is what happened once I let it.
The setup # There is a CLAUDE.md file at the root of my vault. It is twenty-nine lines long, and the last three of them say this:
This is a history of three years - from bookmarking ChatGPT articles to an AI-managed knowledge base; with a certification in the middle and one embarrassing self-assessment along the way.
Spectator to User (2023) # I still remember my first interactions with ChatGPT (via the web, using the free version) sometime in early 2023. I wanted to have an episode guide of an excellent HBO series called Patria, a story of a Basque family during the ETA years. So I asked for it, and what ChatGPT produced was completely, utterly, unbelievably hallucinated: 12 episodes (the series has 8), a made-up story, randomly assigned actors, two thirds of them non-existent. One huge hallucination. I did not keep the transcript, but I remember it clearly. It was very disappointing, and my conclusion was: well, this is fun, but useless.