Field Notes
A weekly dispatch from the edges of the knowledge graph.
Week of
Mar 30–Apr 5, 2026
27 sources / 113 insights
Five things worth knowing
> Eric Siu describes 4 months of running an AI-native company architecture inspired by Jack Dorsey and Roelof Botha's "From Hierarchy to Intelligence" essay. Maps Dorsey's four layers (Capabilities, World Model, Intelligence Layer, Surfaces) to actual implementation at Single Grain using Mac Mini +
> Comprehensive guide to Claude Cowork covering setup, workspace structuring, automation stack (skills, plugins, scheduled tasks, dispatch mode), MCP connectors (Zapier connects 8,000+ apps), browser use, computer use, token management, and masterclass workflows (morning brief, content repurposer, f
> Nyk describes a 3-layer memory architecture for Claude + Obsidian that solves context amnesia: Layer 1 (Session Memory via CLAUDE.md + auto-memory directory), Layer 2 (Knowledge Graph via Obsidian vault + MCP bridge with smart-connections and qmd), Layer 3 (Ingestion Pipeline via brain-ingest tool
> Nick Spisak provides a detailed implementation guide for Karpathy's LLM knowledge base approach. Covers the exact folder structure (raw/, wiki/, outputs/), schema file (CLAUDE.md), agent-browser for automated web scraping, wiki compilation workflow, Q&A against accumulated knowledge, and monthly h
> Ashu Garg (Foundation Capital) and partner Jaya Gupta argue that decision traces — structured records of why decisions were made, not just what was decided — are the enterprise equivalent of B2C behavioral data. AI agents create decision checkpoints automatically: every time a human edits an agent