Member-only story
An LLM Wiki for the Rest of Us
Not a member? Read this story for free here.
Keep the wiki your team edits. Let the model compile everything else.
Ask Claude a question about a folder containing fifty documents, and it reads them. Ask the same question tomorrow, and it will read them again. It finds the relevant passages, weighs them against each other, notices that two disagree, selects one, and writes an answer. Then, it forgets all of that.
Most of us work with documents this way. We upload, ask, get an answer, and repeat. On a larger scale, this process is called RAG (retrieval augmented generation): retrieve fragments, stitch them together, and discard the work. It works. But nothing accumulates. Andrej Karpathy summed it up in one line: “The LLM is rediscovering knowledge from scratch with every question.”
In April, he published an alternative: a short document he calls an “idea file.” You paste it into your coding agent, and the agent builds the rest with you. The idea is to let the model read each source once when it arrives and write what it learned into a wiki. This creates entity pages, concept pages, cross-references, and notes wherever a new source contradicts an old one. When you ask a question later, the model reads the wiki, which reflects every source, instead of reading the sources themselves.










