Vibe Coding Session: Git, GitHub, Permissions, and What Heidi's CEO Builds Himself

Offcall hosted the fourth installment of its physician vibe coding webinar series, and this one pushed past the fundamentals into intermediate and advanced territory.

Dr. Graham Walker and Dr. Michael Hobbs walked a live physician audience through Git and GitHub, the different modes for directing an AI coding tool, and the permission settings that determine how much autonomy you hand over. They were joined by Dr. Tom Kelly, a vascular surgeon and CEO of Heidi Health who still vibe codes features into Heidi's production codebase, and Dr. Liam McCoy, a neurology resident working with the Stanford and Harvard Arise research group, who previewed new benchmark data on how AI models perform across clinical tasks.

Key Webinar Moments

• Why Graham has never used "dangerously skip permissions" mode, and what it actually does
• Heidi Health's CEO explains how he personally shipped 120 pull requests into the company's production product
• A first look at the Arise group's new medical AI benchmark suite, launching the following Monday
• Five physician-built tools demoed live, including an AI-generated cardiology podcast and a communication skills coach
• The plain-language explanation of Git and GitHub that finally makes the terminology click

View the Slide Deck

Git and GitHub, Without the Jargon

The session opened with the vocabulary clinicians need before touching a terminal. Michael Hobbs framed Git as version control: a way to save your work incrementally so that if something breaks, you can go back to an earlier point rather than losing everything. Graham Walker compared it to 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.

A few terms came up repeatedly. A repository, or repo, is the project folder. A commit saves your work at a checkpoint. A branch lets you build a new feature without touching the working version. A pull request, confusingly named since you are technically pushing changes, is how you ask someone to merge your branch into the main codebase. Graham noted that he no longer types Git commands manually. He simply tells Claude to commit and push the changes for him.

Plan Mode, Goal Mode, and Knowing When to Walk Away

Michael demonstrated the difference between working step by step and handing off a task entirely. Plan mode is collaborative: you and the AI think through what you're building together before any code gets written, and the tool checks in at each decision point. Goal mode is a handoff. You describe the outcome you want, tell the AI to run with it, and step away until it's done.

Graham compared manual, step-by-step approval to trusting a chief resident who is a month from graduating: reliable, but you still want to see what they're doing. Auto-approval modes go further, letting the AI run more decisions on its own. At the far end sits a setting literally called "dangerously skip permissions" in Claude Code and "YOLO mode" in OpenAI Codex, which removes guardrails entirely. Graham was direct about it: he has never used that mode and does not plan to, because the consequences of an unsupervised mistake, like a deleted hard drive, are severe enough that the convenience isn't worth it.

Heidi's CEO Is Still Shipping Code

Dr. Tom Kelly, CEO of Heidi Health, joined to describe how vibe coding changed once frontier models became capable enough to work inside large, production codebases with thousands of files and many engineers. He said he personally shipped around 120 pull requests into Heidi's evidence product between January and February, ranging from small styling fixes to a CPD tracking feature he built largely on his own before engineers refined the backend.

Kelly's advice to physicians with an idea and a working prototype but no path to further development: find a small community willing to try it, listen to their feedback, and let it spread by word of mouth before worrying about turning it into a business. He credited his clinical background with giving him instant credibility when pitching early versions of what became Heidi, and said the hardest part was never the idea but figuring out whether an idea was strong enough to justify leaving clinical practice to pursue it full time.

A New Benchmark for Medical AI

Dr. Liam McCoy previewed research from the Stanford and Harvard Arise group, launching publicly the following week: a centralized benchmark suite built to keep pace with how quickly models improve. The suite spans ten benchmark sets, including cases drawn from the New England Journal's clinical pathologic case archive, chest x-ray interpretation, dermatology, agentic reasoning tasks, and a harm-avoidance benchmark.

McCoy's core finding was that no single model dominates across every category. Google's models led on imaging, Claude performed best on agentic execution, and clinical-specific products from partners like GlassHealth, Ambos, and Doximity consistently outperformed general-purpose chatbots on clinical tasks, validating the value of purpose-built clinical AI rather than a single general model. Models also outperformed humans on the harm-avoidance benchmark, but humans still outperformed models on radiology, dermatology, and script concordance testing, a measure of how well a clinician updates their answer under uncertainty.

Physicians Showed What They Built

The session closed with a rapid-fire showcase of tools built by physicians in the Offcall community:

• Nikhil Shah, a nephrologist, built Audio Scholar, which turns a chosen medical specialty's recent literature into a ten-minute weekly podcast with direct links back to the original papers.
• Mike Logan, an emergency medicine physician, built Physician Builders, a sandbox where clinicians pitch tool ideas, get matched with a builder, and iterate with feedback from a global physician community.
• Rishad Usmani, a family physician, built Primary AI, a clinical decision support tool with citation verification and localized guideline and antibiotic resistance data.
• Josh Baron, an emergency medicine physician, built Claire, an AI triage assistant that gathers clinical information from on-call requests and delivers a decision-ready summary to the physician.
• Aaron Reinke built Whetstone Health, a platform that analyzes patient encounters against evidence-based communication frameworks and offers targeted training exercises for CME credit.

Resources and Links

• Michael Hobbs' getting-started guide: ai101.health
• Stanford Arise / HAILS conference: https://arise-ai.org/hails mention Offcall at registration for a discount
• Install Michael's skill with: curl -fsSL https://ai101.health/install | bash

This webinar was sponsored by Evidently.

Follow Offcall on LinkedIn to be notified when registration opens for the next session in the series.