AI Tools for Developers

This is the largest section of the site, and the one closest to my day job. I build ERP and CRM systems in React and TypeScript, and I use AI coding tools on that work every day. Everything here starts from that: a tool I actually put on a real codebase, not a feature list I read about.

A lot of the writing covers model choice, because that is the decision developers make most often and get the least useful help with. Which Claude model to reach for on a given task, what the pricing tiers really cost once you account for retries, when escalating to a bigger model is cheaper than three cheap attempts, and what happens when a model refuses or falls back. I publish the reasoning and the limits of my testing rather than a benchmark table you cannot reproduce.

The rest covers the tooling layer around the assistant: MCP servers and what they are genuinely useful for in frontend work, Cursor and the open-source alternatives to it, Google Antigravity, and the failure modes that come with generated code — hallucinated package names, review blind spots, and debugging code you did not write. There are also straight technical explainers, like the Rolldown change in Vite 8, where the topic is worth understanding on its own terms.

Two things I try to hold to. I say when something is outside what I have tested — my work is TypeScript, React, Node and Postgres, and conclusions from that stack do not automatically transfer. And I date things clearly, because this area moves fast enough that a guide written three months ago can be naming models that no longer exist.