I write the way I build — slowly, carefully, with a preference for the specific over the grand. Most of what lands here is about the LLM transition: where it actually works, where it doesn't, and how to tell the difference. Essays published on Medium.
Every non-deterministic step in an agent workflow deserves a deterministic gate. The case for treating context selection as the sixth gate — and why nobody's building it yet.
Why the command line — not the chat window — is where AI-native development is actually happening. A walk through the tools, the workflows, and the reasons the terminal keeps winning.
A practical case for bounded context: what to hide from the model, what to show it, and why the default of "paste the whole repo" is the wrong one.
Treating specifications as formal contracts changes what LLMs can do for you. Notes from building specBuilder, speccritic, and plancritic.
What happens when multiple frontier models design an editor together — and the surprising, specific tool that falls out the other side.
On using Claude to take on the projects I would never have started alone — and the small set of habits that made the work actually ship.
A walkthrough of building an MCP server for LLM-driven code review — schema design, tool wiring, and the real-world gotchas the docs don't cover.
A distilled set of patterns for using Claude as a Go pair programmer — what to delegate, what to own yourself, and where the handoff goes wrong.
A lab notebook from the first wave of AI-native development — what Claude could actually build, what it couldn't, and what the end-to-end loop looked like.
If humans aren't the primary readers of the code, the language should know. A thought experiment that got less theoretical than I expected.
A from-first-principles walkthrough of MCP — the protocol, the transport, and what it actually takes to stand up a working server in Go.