The deciding part
Writing the code stopped being the hard part a while ago. Deciding what to write, and what you'll still want to own in two years, didn't get easier. AI arguably made it worse, because a bad idea now gets built just as fast as a good one.
Four years on systems that global operations teams run on teaches you that a bad call doesn't show up in the sprint review. It shows up about eighteen months later, as the service everyone quietly routes around. So I've got into the habit of asking what a decision will cost before I make it. And I've kept widening what I'll take on: backend first, then the front end, now the infrastructure. The problems don't respect those boundaries anyway.
What I built for myself
Most people I know use AI as a slightly better autocomplete. I wired mine into the systems that actually hold the context. It runs over MCP against Jira, Confluence and my Obsidian vault, so when I point it at a ticket it already has the story, the technical docs behind it, and somewhere to put the plan it comes back with. The plan lands in the vault and I work from there.
Nobody asked me to build it. It bothered me that it kept guessing at context I could just hand it.
