What's going up, week by week.

Choosing between RAG, MCP, and both.

The RAG-or-MCP question collapses into three architectural questions about what your system reads, acts on, and how it answers. Five worked use cases show what the picks look like.

MCP vs RAG.

The comparison is misframed. MCP is a tool-call protocol; RAG is a retrieval pattern. Most real systems run both, at different layers of the stack.

How to build a custom MCP in your app.

An opinionated playbook for the in-app HTTP MCP pattern. Build it once inside your existing TypeScript app, and it pays back twice: at the keyboard interactively, and on a schedule via Routines.

Anthropic ships Routines.

Claude Code can now run scheduled agents on Anthropic's infrastructure with your MCPs and skills wired in. The piece that was missing for end-to-end automation without a separate cron worker.

The Charlie Hills LinkedIn stack.

One voice profile, eighteen skills, every channel downstream of a single newsletter.

How to build an AI agent.

Why most agent projects fail at step one, and what the eight-step process actually buys you when you take it seriously.

The CLI agent wars settle into a pattern.

Claude Code, Codex, Aider, and Cursor's terminal mode have converged on the same shape. The convergence tells us what is coming next.

Lead with the stack.

State your framework, runtime, and hard constraints in the first line of the prompt. Do not bury them.

Software 3.0.

via Andrej Karpathy

Andrej Karpathy reframes the last three decades of programming as three software paradigms, and puts vibe coding in a longer arc.

Building effective agents.

Anthropic's quiet thesis. Most AI features do not need an agent, and the ones that do follow a smaller pattern than the framework discourse suggests.