> review the git commit history and from that information create a skill that will allow me to create idiomatic git commit messages for this repository

Semantic HTML, proper labels, ARIA roles, meaningful alt text — the things that let screen readers parse a UI are the same things that let AI agents navigate and test it. A `<div onclick>` confuses both. 1/2

The code smells we learned to spot don't apply to AI-generated code. 1/4

These articles are related: doll.leaflet.pub/3mifom65id22r Thanks all, very interesting work Cc: @dollspace.gay, @ai4geo.bsky.social

Three patterns in AI-generated code worth a closer look in review: 1/4

A colleague told me they were worried their skills were getting rusty from leaning on AI. Two things came out of that conversation. 1/3

When we started using agents across our SDLC, the first challenge was figuring out where a model actually helped and where it made things worse. 1/3

AI-generated React components rarely get accessibility right out of the box: click handlers on divs instead of buttons, no keyboard navigation. 1/3

aicoding.leaflet.pub/3miwhqqvwxc2... More great stuff from @chadfowler.com

@dollspace.gay knows this stuff. Its recommendation is going to the top of my todo:review list

When most of a PR is AI-generated, skimming for style and logic isn't enough. We run reviews through a different model — cold session, no prior context, no memory of the choices that led here. No reason to defend them.

Something that’s become a habit on our team: we write the tests before we let the agent write the code The tests end up being a gate the agent codes against. And we get a built-in check that what it produced actually matches what we asked for

We're trying to figure out the rigour involved in getting to a state where humans neither read or write code