Week 14, 2026 Claude Skills, Code Internals and Buddhism

What I learned

The Claude Code source code leaked via an npm map file this week, exposing internal structure and prompt-handling logic
I believe this level of transparency, even if accidental, accelerates the arms race in agentic coding tools. The developer tooling space is about to get very competitive

Learned about Anthropic’s Agent Skills as an open standard this week. This is the most significant Claude platform release since MCP and it shifts Claude from a chatbot toward a composable autonomous agent runtime.
I read Skills Explained which clarifies how Skills differ from prompts, Projects, MCP servers and subagents. Skills are reusable, deployable units of agent capability
Extending Claude’s capabilities with Skills and MCP explains how the two complement each other: MCP for data access, Skills for encoded workflows
Building Agents with Skills and Improving skill-creator round out the launch with practical guidance on measuring and refining agent skill performance
Anthropic published claude-plugins-official with a skill-creator plugin. The open standard positioning is smart: it invites community contribution while keeping Anthropic at the center
Skills was launched back in October 2025. I have been running almost six months late. I knew skills are like a specific system prompt in a markdown file, but that I used to come across bits and pieces. When I started to learn this deliberately to teach to others, I realized how powerful and flexible this system is.

Markdown Engineering’s Claude Code Source Deep Dive is the best structured series I have found for understanding Claude Code internals. Lesson 3 covers the Skills system and Lesson 50 does a full architecture overview
I read Anatomy of the .claude/ Folder by Avi Chawla. A clear breakdown of how CLAUDE.md, commands, skills and memory files are structured
I discovered travisvn/awesome-claude-skills, a curated list of Claude Skills and resources for customising Claude Code workflows
sickn33/antigravity-awesome-skills claims 1340+ agentic skills for Claude Code, Cursor, Codex CLI and more. Overwhelming in scope but a useful reference for discovering what the community is building
shanraisshan/claude-code-best-practice and anthropics/claude-cookbooks are useful companion references for practical Claude Code usage

I explored CrewAI’s documentation and architecture to understand how it structures agent workflows
I discovered CatchMe, described as the missing layer of personal agent. It attempts to build a persistent, personalised context layer on top of existing agents. Still early but the direction is right
I watched Andrej Karpathy’s “Skill Issue” talk on Code Agents, AutoResearch and the Loopy Era of AI. In my view this is required watching for anyone building in the agent space
I spent time with the Claude Cowork tutorial by Design Lead Jenny Wen. 40 minutes of structured walkthroughs that clarified how Cowork schedules and chains agent tasks
Claire Vo’s talk on how OpenClaw’s AI agents run her business, family and life is one of the most honest and practical accounts of actually living with autonomous agents I have come across
I also watched the Claude Cowork Full Course for Beginners which is a solid on-ramp for teams just starting with multi-agent orchestration
I explored Factory.ai’s documentation including their Vibe Coding Skill and Droid VPS setup guides. Factory is positioning itself as the developer-first alternative to Claude Code for agent-driven dev workflows.

I explored Slack Engineering’s post on Prometheus-based network probing for HTTP/3. Their internal Blackbox Exporter did not support QUIC natively, so they extended it for HTTP/3 readiness checks across hundreds of thousands of endpoints. A well-written case study on practical OTel and Prometheus work
I spent time deep in Arize LLM evaluators and online task tracing for production GenAI systems at work
I explored Temporal’s Replay 2026 conference content on durable execution for building reliable agentic systems and started the Temporal 101 with Python course

I was fascinated by Webminal’s story: 15 years, one server, 8GB RAM and 500k users. No Kubernetes, no microservices, no auto-scaling. It survived a datacenter fire in 2021 and multiple power outages in the Netherlands. This is a masterclass in resilience through simplicity
I read Martin Lysk’s post on SQLite on Git and his approach to managing blog images with Excalidraw frame exports

I explored Simon Willison’s browser-based HTML tools, particularly the GIF optimizer, SVG renderer and terminal-to-HTML converter
I read Simon Willison’s note on a Python Vulnerability Lookup tool he built as a single-page HTML app powered by Pyodide. The pattern of shipping useful security tooling as a static page is underrated
Google released Gemma 4 open models. The HN thread had good technical depth on the architecture changes. The open-weights race is still very much on

Axios was compromised on NPM with malicious versions dropping a remote access trojan. Another reminder that unpinned dependencies in production are a liability
I came across the story of Copilot editing an ad into a developer’s PR. If this is real, it is one of the most blatant examples of AI tools betraying user trust I have seen.