A running list of things I bookmarked and found genuinely worth reading (or watching) this week. Heavy on Claude Code and AI agents — that's where my attention has been.
Claude Code & The .claude/ Folder
Lessons from Building Claude Code: How We Use Skills — Thariq, Anthropic
A firsthand account from someone who works on Claude Code at Anthropic. Covers how the team actually uses Skills internally, what patterns they found useful, and some of the design decisions that went into making Claude Code work well for real engineering tasks. Worth reading if you're building skills or workflows on top of Claude Code.
Anatomy of the .claude/ Folder — Akshay Pachaar
The .claude/ directory is Claude Code's control center — it's where you configure how the agent behaves in your project. This article is a complete reference: CLAUDE.md, custom commands, modular rules, skills, subagents, and permissions. One of those posts you'll want to keep open as a reference while setting things up.
How Claude Code Actually Works — Santi
Claude Code isn't a glorified chat interface. It's an agent that runs in an autonomous loop until it completes a goal. This piece breaks down the components of that loop — tools, context, memory, hooks, skills, subagents, MCP — and explains why understanding them changes the way you use it. Good mental model if you're still treating it like a prompt box.
Karpathy's Autoresearch
The Scientist That Never Sleeps — Anish Moonka
Andrej Karpathy released Autoresearch: 3 files, 630 lines of MIT-licensed code. It's an agent that runs the scientific loop autonomously — hypothesis → experiment → result → iterate — while you sleep. In two days it ran ~700 experiments and found 20 genuine improvements that Karpathy had missed in two decades of experience. That last part is worth sitting with for a moment.
Autoresearch: From 41% to 92% with Claude Code — Aakash Gupta
Aakash Gupta (a PM, not an ML researcher) pointed Karpathy's Autoresearch at a benchmark using a Claude Code skill, went from 41% to 92% in 4 rounds overnight, and then wrote up a guide with 6 use cases and 10 evaluation templates for non-engineers. Good proof that this kind of tooling is more accessible than it looks.
No Priors Podcast with Karpathy — Sarah Guo
Sarah Guo's hour-long conversation with Karpathy on the No Priors podcast. They cover the phase shift happening in software engineering, AI psychosis, the thinking behind Autoresearch, and what happens to the job market for developers. One of the better long-form Karpathy conversations out there.
AI Agent Design & Patterns
Developer's Guide to AI Agent Protocols — Google Developers
Stop writing custom glue code for every tool, API, and frontend. Google walks through a B2B agent they built using the Google Agent Development Kit that handles the full stack with 6 open standards. The core argument: the future of agents is interoperable, and we should be building toward that now.
10x Your Productivity with Coding Agents — Matt Shumer, HyperWriteAI
Most people use Claude Code wrong: one thread per task, single agent, no memory. Shumer lays out the correct setup — a long-lived orchestrator thread that delegates to specialized subagents. It's the kind of architectural pattern that seems obvious once you see it, but isn't intuitive when you're starting out.
AI Agents 101 — Free 58-Minute Masterclass — Greg Isenberg
If you're trying to explain AI agents to someone who's not technical (or if you want a clean mental model yourself), this is the one to send. Covers the core loop, how MCP connects tools to agents, why context beats prompts, and how Claude Skills fit into the picture. No prior knowledge required.
Personal AI Systems
Obsidian + Claude Code = Your Own JARVIS — Cyril-DeFi
A walkthrough of connecting Obsidian to Claude Code to build a personal AI system that knows your notes, tasks, calendar, and ideas. The end result is an assistant with full context about your work and life — rather than a general-purpose model that starts from zero every conversation. Practical setup guide.
obsidian-web-mcp — Jim Prosser
A remote MCP server that lets Claude interact with your Obsidian vault through a web interface. Useful if you want to query or update your notes from an AI agent without setting up local infrastructure. Clean, minimal, open-source.
AI-Native Design
Guide: AI-Native Design with Paper — TK Kong (ex-Ramp)
A practical guide to designing products that are built around AI from the start — not retrofitted. TK Kong (formerly at Ramp) walks through "Paper," a design system concept built natively for AI interfaces. Changes how you think about UI components, data flow, and interaction patterns when the model is doing real work.
Tools Worth Knowing
Top 50 Claude Skills & GitHub Repos for AI — darkzodchi
After scanning 1,000+ repos and testing 200+ skills, the author picked the 90 tools that actually matter: Claude skills, MCP servers, and GitHub repos with no filler. A good list to bookmark for the next time you're setting up a new workflow.
OpenClaw: Complete Guide — Alex Finn OpenClaw: Definitive Masterclass — Greg Isenberg + Moritz Kremb
Two separate takes on OpenClaw, the open-source personal AI assistant. Alex Finn covers the full setup, and Greg Isenberg + Moritz Kremb do a step-by-step 11-part guide for going from "I installed it" to "it actually works for me." If you've tried OpenClaw and bounced off it, the Greg/Moritz walkthrough in particular fills the gap.
AI Agents Building Real Products
Larry, My AI Agent, Generates Millions of Views — Oliver Henry
Oliver Henry built an AI agent called Larry that autonomously creates TikTok slideshows. The original article got 7 million views. He's now launched LarryLoop — the same tech packaged as a web app with no server setup or terminal required. An interesting case study in what happens when you productize an agent that was working for you personally.
Not AI
Malcolm Gladwell: Outliers — Al Panda (via Brad)
A summary and discussion of the core argument in Outliers: extraordinary success isn't about individual talent or effort alone — it's a combination of systemic factors, timing, and circumstances that most success narratives deliberately leave out. Good reminder to zoom out when thinking about what actually drives outcomes.
That's the week. Most of my reading time went into Claude Code and the Karpathy orbit — it feels like a lot of pieces are starting to connect in interesting ways.