8 posts filed under AI, newest first.
Most of what makes Claude Code work for me isn't prompting, it's a few files I wrote once: a plan-mode habit, a /simplify pass welded onto /commit, a hook that won't let me hand-write a commit, and a CLAUDE.md that's mostly scar tissue.
I came back to a project after 14 months and found two notes I'd left myself, in the same file, in the same voice, and one of them saved me 40 minutes and the other cost me an hour and a half.
Eight months into their first job, they fixed in 40 minutes a bug that would have taken me two days at that stage. Then they couldn't tell me why the fix worked.
Three Claude skills cover the whole life of a Shopify client repo at work, and after writing all three I'm fairly sure the steps are the least valuable thing in them. The value is the four or five places where the runbook is told to refuse.
The alignment people and the stay-ahead-of-China people are both right. Inconvenient. I still land on keeping our foot down while building the safeguards into the same motion.
Three people in one standup gave three different multiples for how much quicker they've got, none of them had measured anything, and I sat there working out whether to say that I genuinely don't know if I'm faster at all.
I asked for test coverage on a module that didn't have any, got 14 tests back, and all 14 passed, and about half of them could not have failed under any circumstances whatsoever.
A real component, plausible output in nine seconds, and I kept about a third. The two thirds I threw away were wrong in a specific and predictable way.