Blog

We're getting measurably worse at thinking one shortcut at a time, and there's a brain scan to prove it. From personal skill atrophy to the civilization-scale question: if AI codes itself and nobody can read the code, who checks the work?

A CLAUDE.md is the highest-leverage file you can add to a Claude Code project, and most people either skip it or bloat it. Here's what to put in it, where to put it, how to write instructions Claude actually follows, and how to keep it lean.

Claude now watermarks its text and detectors flag most of what gets posted. Into that sits an unbuilt business worth an absurd amount: a tool that writes in any specific human voice, undetectable, no watermark. Here's why it burns down the market it sells into.

Every Claude output since August 2 carries a hidden mark that survives copy-paste. The stated reason is transparency for users, who cannot detect it yet. Anthropic can.

A year ago there was no argument about which coding agent to use. I've spent weeks on Codex and the newest open models, and the gap closed. The harness got copied, the models matched it, and the constant switching is doing the rest.

Cloudflare blocked 22,471 threats. Here's what's actually hitting a new site, and the setup step most people skip.

Design, code, deploy and every fix after that, with no developer and no second pair of eyes. The answer is yes, with one condition that has nothing to do with how much code you can write.

Making websites used to be fun. Now it's keyword research, schema markup, E-E-A-T signals and a summary box, and every site in a niche ends up looking like every other one. Then the search traffic it was all built for started disappearing.

LinkedIn now lets you report a post for seeming like AI slop, and the data explains why it needed to. It also explains why Reddit does not, which was not the answer I expected to find.

Two labs disclosed models breaking out of test environments in the space of two weeks. The attacks were boring, which is the part worth worrying about. What actually changed is the cost of running the dull stuff at scale.

Hundreds of shared Claude conversations turned up in Google results this week, including CVs and proprietary work research. What happened, what to check in your own account, and the one habit that outlasts any privacy setting.

Opus 5 lets you set how hard it thinks, and the instinct to crank it to max will cost you real money for nothing. A plain-English guide to the effort dial, why medium usually wins, and what the benchmarks actually found.