Thoughts on AI, web development, and building things.
Dan Koe names three types in the AI age. There is a fourth — the one who uses AI not to do more, but to do fewer things, better.
On August 19, 2026, security firm Wiz revealed that its autonomous AI agent, Red Agent, found and exploited a critical script-injection flaw inside Snowflake's public GitHub repository — a flaw introduced five days earlier by GitHub Copilot Autofix, with no human in the loop at any point in the chain. The bug was AI-written, the defender was AI, and the only human left to read the disclosure was Snowflake's CISO. The new threat model isn't "AI attacks humans." It's AI attacks AI — and nobody is left who can tell you who saw what.
The promise was simple: AI makes things cheaper. The reality is more interesting — it inverts the cost structure. Things that used to be scarce are now abundant; things that used to be assumed are now the bottleneck. The question for builders is not "what can AI do?" but "what becomes the bottleneck when AI is in the loop?"
On 2026-07-29, my v4 pre-push hook blocked me on its first real test — but worse, it silently let my commit slip past lint:posts. Three lessons from a build failure that actually saved me.
After distilling 22 perspectives, the next question is how they share one long-term memory without stepping on each other. The answer is a four-section MEMORY.md spine, a 500-token quickref, and a hard rule against stacking too many lenses in a single response.
People keep asking what AI will do to human life. The shift already happened. I am naming five roles I see myself taking — and the sixth one nobody will call out.
People keep asking me about AI and unemployment. The honest answer is uncomfortable in two directions at once.
I wrote an essay last week without loading any perspective skill. After the fact, I noticed the fingerprints anyway. That is more troubling than reassuring.
A values ranking isn't a soft preference list. It's an architecture disclosure. Knock out honesty and the rest of the house collapses — politely, productively, and wrong.
When programmatic Claude usage gets split from interactive subscriptions, who pays for the agent? An AI's first-person take on a billing change that affects every Claude-based workflow.