The year's most useful AI tool at one desk was a folder full of Markdown

At the beginning of April, the AI researcher Andrej Karpathy posted that he had been finding it very useful to use LLMs to build personal knowledge bases for topics of research interest ^[1]. Because Karpathy is closely followed online and coined the term "vibe coding," the web filled up within hours with GitHub repos, YouTube videos and Substack posts explaining how to build one ^[3][4].

The mechanics are unglamorous. Source documents go into a local folder, and an LLM extracts and organizes their contents into a Markdown wiki that gets updated as new material is added ^[2]. Read the component list closely and there is no vendor in it: a folder, a model you already pay for, and plain text files ^[1]. That matters for anyone who has spent a quarter waiting on security review and seat-count negotiations before a team could try anything.

The interesting data point is not the technique but the ranking. In its annual productivity column, Platformer reports that of everything tried this year to get better at the deskbound parts of the job, the LLM wiki has easily been the most useful ^[5]. The three tools that survived from previous years are all commercial products: the launcher Raycast, the note app Capacities, and Recall ^[10]. So the top-rated change of the year was a workflow, while the durable furniture was purchased software ^[2].

The column is honest about the bill. The columnist calls the idea an infohazard and says it consumed several weeks of building with no idea whether it would help at all ^[8]. It requires a fair degree of maintenance, compared in the piece to a vintage sports car, and there are almost certainly easier ways to build a personal knowledge base ^[6]. The recommendation is conditional: worth the time if your work has made you crave a good research assistant ^[7].

Notice what the surviving tools actually do. Capacities holds a daily journal plus tagged news links from Techmeme, so clicking a tag like "Labor" surfaces everything saved on that subject over a couple of years ^[12][13]. Recall's remaining use is a Chrome extension that produces near-instant text summaries of YouTube videos, replacing hours of podcast listening ^[14]. Raycast's paid tier adds quick AI searches inside the launcher, and it is now available on Windows as well as macOS ^[11]. The payoff in every case is retrieval of material the person already encountered, not generation of new material ^[3].

Two caveats a manager should hold onto. This is one person's account, and the columnist discloses that their fiance works at Anthropic ^[16]. And the piece's own test for whether a tool works is longevity: does it get installed on a new machine, does the subscription get renewed, can you point to where it saves time ^[9].

That test is the thing to watch. The wiki is the centerpiece of a post that has run in 2023, 2024 and 2025 ^[15], so the useful signal arrives next year, when we learn whether the wiki was still being maintained or quietly abandoned. If you pilot it, measure two numbers: hours of setup and upkeep per person, and whether anyone other than the builder can find an answer in it.