Phase One, or: This Is Not Vibe Coding

A complete course for developers who are serious about making AI part of how they actually work.
The code map and project dump come back with the right answer: Moebius is a small Elixir library for querying Postgres, not an ORM, and it still compiles, with a few bugs and an upgrade plan. Before I act on it, a word about vibe coding, because a lot of what we do these days smells like it. Vibe coding is sitting with a CLI and riffing out an app with no tests, just to see if it works. I have a test suite here, which means I have accountability, and I'm not going to shotgun changes without it. The difference between this and vibe coding is accountability, and we'll keep coming back to that.
Phase one is to get it compiling and running the tests locally without touching the driver. I ask for a plan.md in the root so I can see the tasks it's working against, and I'm explicit about what not to do, because a blanket "go ahead" to Claude with bypass permissions on means it will do a lot of things.
OK, let's do phase one, which is to upgrade it. I want to have a plan.md here in the root, so I understand the tasks that you're working against. And then let's go ahead and execute them. I don't want to deal with swapping out the driver just yet, so don't touch that.
Yes, I run with bypass permissions. The default is for Claude to ask before every file write, which gets so annoying that you hit OK without reading, and that's worse than not asking. The more you work with AI the more you trust it, and I'm doing this on main with the driver swap coming in a branch I can delete. When it finishes, Moebius builds and tests cleanly on Elixir 1.20 and OTP 29, driver untouched, 104 tests passing (plus three flaky ones I believe live down in Postgrex). That would have been four or five hours by hand. It would also be a 22-commit PR that no human reviewer at your company wants to see, which raises a question I want you to sit with: what do we do now that the tool reviews code better than we can?
I commit and push through the GitHub CLI, which is authenticated as me, so Claude can create repos, issues, changelogs and readme updates on my behalf.
Commit these changes, add all the new files, and then push everything to GitHub.
Before signing off I look at the plan the agent worked against, and at the CLAUDE.md it wrote: the main instruction set for Claude, telling it where the code map is, how to run the tests, fetch dependencies and check code quality. This is one of the most important files in your project. We're slowly turning Moebius into an AI aware project, and next we do the big work.
A complete course for developers who are serious about making AI part of how they actually work.