Git for Physicians: A Guide to AI Vibe Coding Safely

Demystifying Version Control for Clinicians

For physicians stepping into the world of "vibe coding" and AI app development, the barrier to entry is rarely a lack of clinical ideas. More often, the hurdle is the intimidating software development terminology. As Dr. Michael Hobbs noted during a recent Offcall webinar, going into a computer's command terminal is "like going to a foreign country. You want to learn a few key phrases."

At its core, the concept of Git is simply version control. As Dr. Hobbs explained, it is "a way to save your work incrementally so that if something breaks, you can go back to an earlier point rather than losing everything." It is a foundational safety net for anyone writing code, ensuring that hours of hard work are never accidentally overwritten or destroyed.

Dr. Graham Walker likened the process to a practice many seasoned clinicians know all too well from their academic days:

"...the old habit of saving a paper to a floppy disk or emailing it to yourself as backup, except Git does it automatically and shows you exactly what changed between versions."

Key Vocabulary You Need to Know

Before you touch a terminal or open an AI coding environment, understanding the basic lexicon will make the process infinitely less daunting.

The best part of modern vibe coding? You do not have to memorize complex command-line codes. Dr. Walker admitted his own shortcut during the session: "I don't really do Git commands anymore. I just ask Claude to do it for me... I now say, 'Hey Claude, can you merge, commit, and push this?'"

Choosing Your AI Coding Mode

When directing an AI coding tool, you can dictate exactly how much autonomy the model has. The webinar hosts broke down the primary modes of operation:

• Plan Mode / Ultra Plan Mode: This is a highly collaborative approach. You and the AI think through the architecture together before any code is actually written. The AI acts as a thought partner, checking in at various decision points to ensure it understands your vision.
• Manual Mode: The AI asks for explicit permission for every single step it takes.
• Goal Mode: A complete hand-off. As Dr. Hobbs noted, you give the AI a clear directive and essentially say, "start here, don't come back to me until this is done because this is your goal." You can even ask Claude to write the perfect, highly detailed "Goal Prompt" for itself.

Navigating Permissions and Safety Guardrails

Boundary-setting is vital when working with AI agents that have access to your computer's file system. As you move away from manual approval, you enter the territory of auto-approving commands.

Dr. Walker used a classic medical training analogy to explain this risk spectrum. Auto-approval for basic commands is like trusting a reliable chief resident who is a month away from graduation. You let them run the team and make decisions, but you still keep an ear out.

However, both hosts strongly warned against max-autonomy settings, which are literally named "dangerously skip permissions" in Claude Code and "YOLO (You Only Live Once) mode" in OpenAI Codex. Dr. Walker likened this to letting a brand-new intern completely run the team unsupervised.

"I have never used dangerously skip permissions or YOLO mode... literally never," Dr. Walker stated. Handing over full, unmonitored control to an AI risks disastrous mistakes, such as inadvertently wiping an entire hard drive. For physicians building healthcare tools, maintaining those guardrails is a non-negotiable best practice.