I rolled out Claude Code on my own repos a few months ago and, like most people, the first week was a mix of "this is magic" and "why did it just delete my test fixtures." Below is the short, opinionated playbook I wish someone had handed me on day one.
Day 1 — Don't let it write code yet
Open a sandbox repo, point it at a small bug, and only let it read. The goal of day one is to feel how it budgets context, where it gets confused by your file layout, and how it asks for clarification. You'll save hours later.
Day 2 — Give it a single job per session
Resist the urge to "set up the project and then start the feature." Each session gets one commit-sized goal. If you find yourself saying "and also…", start a new session.
Day 3 — Wire up review, not generation
The first time I trusted Claude to open a PR I regretted it. The first time I trusted it to review a PR I didn't. Start there. A reviewer is bounded, idempotent, and easy to roll back.
Day 4 — Promote the good patterns into commands
Once you find a prompt that actually works for your codebase, save it as a slash command. Treat your .claude/commands folder like a build script you actually own.
Day 5 — Measure, don't vibe
Add a tiny log: how many suggestions did you accept, reject, edit? If your accept rate is below ~40% something is off in how you scope the task. The number is a mirror, not a scoreboard.
Day 6 — Teach it your stack's failure modes
A two-line CLAUDE.md that says "this codebase uses Tabs, not spaces, and never touch migrations/" saves more time than any prompt template.
Day 7 — Review your own work, with it
Open a session and say: "be a senior reviewer of this branch, focus on what I missed." That feedback loop is where the real productivity compounding lives.
That's the loop. None of this is novel — but the discipline of doing it for a full week is. If you try it and a step blows up, drop the failure in the comments; I'm collecting patterns for a follow-up.
Top comments (1)
I'd underline Day 3. One thing I'd add: reviewing AI code is getting harder. The code reads clean now, so the bugs are in the use case nobody scoped, or the scale it falls over at. On home projects I've mostly stopped reading the diff and just test that it works. At work I still read all of it, because AI code is my code and the cost of a mistake is high. But the pace is much slower.