Shipping a small change rarely takes long. Everything around it does. Before the code lands, someone reviews the diff, opens the pull request with the right description, decides who needs to approve it, tells the team it exists, updates the ticket, watches the build, and finally deploys. None of it is hard. All of it is repetitive, easy to skip, and easy to get subtly wrong.

▎

▎ I work on a large codebase and got tired of paying that tax by hand. Instead of writing a longer prompt every time, I turned each step of the ritual into a command the AI runs from the terminal: one to review a branch before the pull request exists, one to open it and classify it, one to hand it off to the team, one to deploy. Each command carries my team's actual rules, not generic best practices, and each one talks to the tools we already use.

▎

▎ This talk is the honest version of that story. I will show the commands running on a real change, then open them up: they are plain text files, and you can read one in a minute. I will also show where it went wrong. My first review command produced findings nobody wanted, and cutting it down mattered more than adding to it. Throwing more AI at the problem did not make the answers better. And a command that tries to do everything ends up doing nothing well.

▎

▎ You will leave knowing how to pick the one step of your own workflow worth automating first, how to write it, and how to tell whether it is actually helping. Nothing here is tied to a specific stack or tracker: the pattern is the same whether your team lives in one tool or six. No prior experience with AI agents required. If you have ever pasted the same instructions into a chat window three days in a row, this is for you.