Honestly, hard to say.
I gave it a ~2 min voice-to-text about my concept for the memo and, after retrieving relevant context via MCP—(1) my core context (claude.md, user.md, memory.md, etc. equvalents), (2) all contexts in my library from nearly a decade of writing about knowledge management, (3) knowledge of the Github repo for the project, and (4) Karpathy's post—it generated about 99% of the post (I changed a link and some formatting). That wouldn't have been possible without years of context, but this post was very much generated from such.
The beauty of the system is that knowledge, clarity, etc. compound.
