$2,800, 21 repos, 24 stars: My first year of vibe coding
1000 commits. Three deployed platforms. One breakout article. And a more honest understanding of what AI-assisted building actually is.
Last September, I sat down on a Sunday afternoon and spent $100 in Claude Code credits over three days. By the end of it, I had a working multi-agent demo AgentInterOp connected to a Vercel deployment, pushed to GitHub, documented with an API spec. I was presenting it at the HL7 FHIR Pittsburgh Connectathon the following week to demonstrate breast cancer screening evaluation via agent-to-agent communication using the A2A protocol.
It worked. The demo ran. Josh Mandel’s language-first interoperability theme was the backdrop. I felt like a builder.
What I Built
Between April 2025 and April 2026, I shipped or vibe-coded into existence multiple things that matter to me.
Key findings:
By commits (top 5):
AgentInterOp (174) - My connectathon sprint paid off in raw output
SmartHealthConnect (158) - Most scaffolding/iteration
fhirquiz (98) - Surprising amount of work here
fhiriq-nextjs (94) - Brand/platform work
FhirMapMaster (84) - Core FHIR tooling
By stars (external validation):
HealthClawGuardrails (13★) - Clear winner, people want AI guardrails
AgentInterOp (3★)
FhirViewMapper (3★)
By usefulness rating:
5/5: HealthClawGuardrails - solves a real, specific problem with production-grade features
4/5: FHIRBuilders, SmartHealthConnect, AgentInterOp - functional but need polish
3/5: FhirMapMaster, FHIRspective, fhirquiz - useful tools, incomplete
2/5: Most FHIR-IQ org repos - early experiments
AgentInterOp started as that Claude Code before the connectathon. It became a multi-agent healthcare platform with dual-protocol support across A2A and MCP, six specialized agents, a scenario engine, Claude AI integration, and 174 commits before I stopped counting. It’s live on Vercel. Nobody uses it as a product, but I’ve demo’d it at two community events and it consistently generates the most interesting hallway conversations I’ve had in the FHIR space in years.
SmartHealthConnect became a full-stack SMART on FHIR patient record application with Epic and Cerner auth, a health dashboard, AI-powered insights, and an MCP server with tools and MCP App integration. 151 commits. An MVP that genuinely works. I used it to pull my own longitudinal health data across providers for the first time and discovered I had better lipid trend visualizations than anything my actual care team had shown me from within Claude interface. That felt significant. And yet, again no users beyond me and my family yet.
FHIRBuilders started as a community platform idea and became a live Next.js application with a Medplum FHIR sandbox, an API explorer, a project showcase, and the OpenClaw generator section. I built it under deadline pressure for an “Agents on FHIR” community meeting demo. It exists, it deploys, and the concept is right. The execution is, to be generous, a strong foundation.
HealthClaw is the most recent and most ambitious an agentic patient-side healthcare assistant that wraps all of the above into a single coherent product vision. It is currently in the beta working phase.
In February of this year, I wrote something in my notes that I want to reproduce here because it’s the most useful thing I can offer anyone considering a similar year: “None of these have users because none of them solve a complete problem for a real person end-to-end.”
That was hard to write then. It’s useful now.
The Tools, Honestly
I used three AI coding environments meaningfully over the year. Each one does something different, and the distinctions matter more than the hype suggests.
Replit is where I started and you go when you need to ship something fast without a care in the world about architecture. The credits don’t go as far it works when you have a easy use case.
Claude Code is where I did the serious FHIR work. Building CQL measures , LDL cholesterol compliance, ADHD medication monitoring and pushing them to my Medplum FHIR server required: sustained, deep context about a complex domain. Claude Code could hold the profile requirements, the value set structure, the denominator exclusion logic, and the resource shapes all at once in a way that meaningfully shortened the implementation gap. Not eliminated it. Shortened it. The IG validator step does not go away. Profile conformance still matters. A CQL measure that produces wrong denominator logic isn’t a syntax error it affects quality reporting. Claude Code understands that..
OpenAI and ChatGPT made appearances early in the year, particularly on AgentInterOp. I’ve since converged almost entirely on Claude Code for FHIR-specific work. The domain depth difference is real.
The Learnings
Here’s the distinction I’ve landed on: vibe coding at its best is an extraordinary learning accelerator and an extraordinary prototyping tool. It is not, by default, a product-shipping machine. The gap between a working demo and something a second person can use and rely on remains large. Vibe coding compresses the distance to the demo, not the distance to the product.
In a domain like building with FHIR in healthcare, this compression has real value. Being able to stand at a connectathon and show a working agent-to-agent breast cancer screening evaluation even a fragile one changes the conversation. Before AI coding agents, I couldn’t have built that in three days. With them, I could. That’s not trivial. But it’s also not a shipped product.
The honest frame for what I built this year is a portfolio of sophisticated proof-of-concepts. Each one demonstrates something technically interesting about what’s possible at the intersection of FHIR, AI agents, and modern tooling. None of them has earned users. Understanding why is the more important question.
What Actually Reached People
The article that got the most traction from everything I produced this year wasn’t code. It was a Substack post.
How I Build My Personal OpenClaw used my own colonoscopy billing dispute, my family’s FHIR records pulled from multiple providers, and the story of building a personal health agent as its narrative spine. It went places none of my repos went, for reasons that are worth understanding.
The demo that got the most live reaction this year wasn’t a GitHub repo. It was the Vibe Coding Happy Hour I hosted in March forty people virtually, four breakout rooms building tools live, with me narrating prompt reasoning out loud and showing failures in real time. The energy in that session was unlike any pull request I’ve ever merged.
This isn’t an accident. It’s pointing at something real about where leverage actually lives in this kind of work.
The code I write is valuable primarily to me as a learning mechanism, as a thinking tool, as a credibility substrate that lets me write and speak from a place of genuine implementation experience. The article and the event are where that substrate becomes useful to other people. I’m not saying don’t build. I’m saying understand the chain of value more clearly than I did at the start of the year.
Why Healthcare Is Different
Vibe coding in healthcare is harder than vibe coding in most domains, and I want to be specific about why.
FHIR as a structured input is actually a gift to AI coding agents. The resource shapes are well-defined, the profiles are documented, the APIs are RESTful and predictable. Compared to most integration work, FHIR is a friendly surface. Claude Code understands FHIR R4 resources at a level that genuinely accelerates implementation.
The problem is what surrounds the FHIR layer. Implementation Guides are long, complex, frequently updated, and their conformance requirements don’t disappear because your AI agent didn’t know about them. A CQL measure that executes without errors can still be wrong producing systematically incorrect population counts because the denominator exclusion logic missed an edge case in the specification. That’s not a hallucination problem. It’s a domain expertise problem. The AI can write the code. It still needs a clinical informaticist to validate the logic.
The same gap shows up on the patient data side. When I pulled my family’s longitudinal health records through SmartHealthConnect, I found what felt like a revelation: I had a true picture of my lipid trends, my kids’ wellness visits, my wife’s care history, assembled from multiple EHRs for the first time. I could reason about starting a statin in a way no single provider visit had made possible. That’s real value. But I also need to be honest: I’m a data guy. I know where the data lives, I know which EHR systems expose SMART on FHIR endpoints, I know how to interpret what comes back. The average patient doesn’t have that context. The technical barrier has compressed, but it hasn’t collapsed.
This is where the “vibe coding is great for frontend ideas, but healthcare backend infrastructure remains notoriously rigid” framing I’ve heard on the podcast comes from. I’ve lived it. The frontend and the demo get you most of the way with AI assistance. The backend conformance, the clinical logic validation, the regulatory compliance layer those still require expertise the tools can’t substitute.
Where This Is Going
I’m convinced we’re one year away from the barrier dropping another order of magnitude. The thing I’ve been building toward an agentic patient-side healthcare assistant that uses FHIR as the data substrate, MCP as the tool protocol, and A2A as the agent communication standard is technically feasible today. The pieces exist. What I’ve spent the year doing is understanding how they fit together, where the real friction is, and what a genuinely useful version of this looks like.
The differentiator going forward is not coding skill. That’s the thesis I’ve staked a year of building on, and I believe it more now than when I started. What used to take a year to build can take days. The question isn’t whether you can build it. It’s whether you understand the problem well enough to know what’s worth building and having a sandbox to iterate with real live customers.
That’s a domain expertise question. In healthcare, it’s also a trust question and a liability question and a data governance question. Those don’t go away because an Agent can scaffold a Next.js app in thirty minutes.
The builders who will matter in healthcare AI are the ones who have both the domain depth to know what a real problem looks like and the technical fluency to build toward it without waiting for a software engineering team to catch up with the idea. The gap between those two things has shrunk dramatically. But it hasn’t closed.
What I’d Do Differently
Build one thing that one person uses every day before building three things that look impressive on a portfolio page.
I knew this going in I wrote about the OpenClaw lesson explicitly. Peter Steinberger didn’t plan his way to 145K stars. He scratched his own itch in one hour, shipped it, and iterated daily in public. I did the opposite: I planned architectures, wrote strategic documents, built interconnected platforms. None of them have a user because none of them solve a single complete problem for a single real person.
The vibe coding tooling makes this failure mode easier to fall into, not harder to avoid. Because the cost of starting a new project is near zero, the temptation to start over rather than push through the hard last mile of actual adoption is constant. I started four projects this year when I should have finished one.
The article reached people because it was finished. It had a beginning, a middle, an end, and a specific thing it was saying. The repos don’t have that. They’re open loops.
Close the loop.
Additional Information
