Review now costs more than writing: two subscriptions pay for the agents. Claude Max covers writing the code, plus everything else they do. cubic reviews every pull request before I merge. This month cubic overtook Max for the first time, on ~60k lines of code in 30 days (most of it the outbound engine: agents find and score prospects and draft the first message, I review and send). Steve Yegge has been saying human code review is over. It mostly is, but the review didn't go away – it moved to an agent, and it's now the bigger bill.

A hook that stopped firing: Claude Code recently changed how it edits files in auto mode (via Bash rather than the Write and Edit tools). My ‘prettier’ hook was triggered on Write|Edit so it stopped working without me knowing and files went out unformatted for days. The fix was moving enforcement down a layer: a pre-commit hook in each repo fires on the commit, whichever tool (or person) made the edit. Anything tied to how the agent works breaks when the agent changes, so connect it to the work instead.

The scratch note: agent replies kept running long, and it turned out the agent had nowhere else to put its working. Reasoning (the thinking block) is wiped between turns, so anything it worked out and didn't write down was gone, and the reply was the only place its instructions gave it to write conclusions. One clause fixed it: detail with no home (conclusions, rejected options, dead-ends) goes in a /tmp scratch note that only the agent reads. A day later a heavy session wrote one unprompted, full of exactly the detail that used to bloat the reply. Extends the reply-length post – full write-up coming. Blog post

"Chat with your data" needs 3 things: a thread this week on whether AI querying your data can be trusted. It can, with: accurate data (almost nobody has it), a canonical layer that standardises your metric definitions, and a skill that tells the agent how to query that layer. Never the raw tables. With those in place anyone asking gets the same answer, however they phrase it. We've just delivered this for a Fidero customer. Once the data is accurate, complete and in the right place, doing things with it is the easy bit.

Facelift without breaking the build: a founder asked how to make a Claude Code app look good mid-build. Give Claude better tools, not better prompts: connect the Stitch MCP, have Claude write a design system from your brand first, then have it prompt Stitch for options and adapt the HTML/CSS back into your site. I get it to lay the options out on a single bake-off page to choose from, and the Chrome MCP verifies every UI change visually. Stitch MCP setup

On my radar

Cursor published an "unslop" skill: a week of AI-speak complaints ("load-bearing", essays for answers, the usual). Cursor's response is a skill listing 31 tells to strip from any writing, puffery through to em-dash overuse. A decent list and quite the admission about default output. GitHub

Everything I own, owned: a security researcher pointed Claude at the firmware of five peripherals on his desk. 13 hours of agent time and 98 prompts later: a plaintext command shell inside his microphone, a webcam whose recording LED he can switch off, and a key light that accepts memory writes from anyone on the wifi. Blog post

Tokens as aluminium: aluminium cost more than gold in 1852, fell 99.9%, and every price drop opened a bigger market than the one it killed. The argument is that tokens are on the same path but ~10x faster, and the step that matters is where the consumer stops being human. A reply from a model router: reasoning is now 50.6% of the tokens they route and no one reads any of it. Tweet