What it is
A personal note by Andrej Karpathy, an AI researcher formerly at OpenAI and Tesla. It is an idea with no product or commercial offer behind it.
The design has three parts. Source documents sit in one folder and are never edited. The assistant writes and maintains a set of plain-text pages from them, one per topic, company or source, with links between pages. A short schema document sets the structure. Three routines run against it: adding a source and updating the pages it touches, answering a question from the pages with citations, and a periodic check for contradictions, stale claims and pages nothing links to.
Two files hold it together: an index of every page by category, and a running log of what was added and asked. Answers cite the page, which cites the source, so a reader can trace any claim back.
Karpathy describes it as deliberately abstract. There is no code or template. The value is the division of labor, where a person curates the sources and asks the questions and the assistant does the maintenance.
What marketing teams should take from it
- The sources are never edited; the assistant maintains the pages. That separation is what lets a reviewer trust a page without re-reading the deck it came from.
- The contradiction check is the part most teams lack. When a release note supersedes an old claim, the pattern surfaces it rather than leaving both in circulation.
- Every answer carries a citation chain, so a reviewer can check the source rather than trust the summary.
- It is written for one person running an assistant against a folder. Permissions, several editors and audit beyond the log are not addressed.
Where to start
Start with the finding 01.2 set for one priority industry. Put the analyst reports, release notes and approved case studies in a source folder nobody edits, and have an assistant write one page per claim, buyer problem and proof point, each dated and citing its source. Read the first twenty pages yourself before anyone else uses them.
Give one person the job of running the contradiction check after each release and retiring stale items. Measure it by the number of stale or unsupported claims caught in review, before and after.
Keep in mind
A pattern, not software. Quality depends on the sources you feed it and on someone reviewing what the assistant writes.