If you use an AI coding agent daily, you have a toolbox of prompts and commands you've never once cleaned up. Here's the refactor, how to spot when yours needs it, and what happens when a tool outgrows the repo.
My terminal has an org chart now. Fable does judgment, Opus ships, Sonnet handles the small stuff, Haiku does the summaries, and the cover art is outsourced.
From ordering Claude around like a code monkey to running real refinement sessions with it. The three things I changed.
The more senior you get, the higher your bar for 'good enough' quietly climbs, until it lands on flawless. But nobody asked for flawless. They asked for something that works.
AI can write the code, the tests, and the commit message. It can't tell you your font is exhausting the reader you forgot to picture. Knowing who the work is good for was always the job, and it still is.
It doesn't tell you the truth. It tells you what you came to hear. The danger was never the AI. It was the access policy built around it, and the context you fed it.