How I AIHow a 91-year-old vibe coded a complex church event management system using Claude and Replit
Claire Vo and John Blackman on a 91-year-old builds church event software via Claude and Replit.
In this episode of How I AI, featuring John Blackman and Claire Vo, How a 91-year-old vibe coded a complex church event management system using Claude and Replit explores a 91-year-old builds church event software via Claude and Replit John Blackman, a 91-year-old church volunteer, describes turning a manual paper-based “impact weekend” registration process into a full web app by using Claude to generate a roadmap, user stories, and requirements, then feeding that into Replit Agents to implement the system.
At a glance
WHAT IT’S REALLY ABOUT
A 91-year-old builds church event software via Claude and Replit
- John Blackman, a 91-year-old church volunteer, describes turning a manual paper-based “impact weekend” registration process into a full web app by using Claude to generate a roadmap, user stories, and requirements, then feeding that into Replit Agents to implement the system.
- The app supports multiple churches and events with role-based admin levels, participant registration via QR codes, service toggles, waivers and signature capture, and printable “impact passports” (PDFs) that attendees can bring to the event.
- It also generates operational reports (demographics for pastoral follow-up, service usage counts, food pantry/lunch ordering, vision center details) and includes an oil-change workflow that looks up oil/filter specs from a VIN using an API.
- They discuss practical realities of agentic coding—fast progress and low cost (hundreds of dollars), but real deployment pain around dev vs production differences, PDF/email attachment issues, and secrets management—plus broader advice about embracing new technology as they did with AutoCAD decades earlier.
IDEAS WORTH REMEMBERING
5 ideasStart with a clear outline, then force the AI to ask questions.
John began with a Word document describing the workflow and explicitly instructed Claude to ask for missing details (“If you need more information… ask me the questions right away”), which produced more complete requirements and fewer gaps during building.
Using Claude as a product manager accelerates agentic coding success.
Claude generated an MVP roadmap, user stories, and non-functional requirements; those artifacts became the “source of truth” that Replit Agents could reliably implement.
Agentic tools can produce “enterprise-like” patterns from community needs.
The app includes multi-level admins, multi-church data separation, approvals for administrators, and reporting for ministry follow-up—capabilities that mirror multi-tenant SaaS design despite originating from a paper process.
Operational reporting is where the real value lands for events.
Beyond “a form,” the system outputs demographics lists for follow-up ministry, service-usage summaries, oil-change shopping lists, and food/vision center prep reports—reducing manual planning and errors.
The dev-to-prod gap is the biggest vibe-coding reality check.
John repeatedly hit issues where features worked in development but failed in production (PDF email attachments, environment variables/OpenAI key reuse), highlighting deployment, caching, and secrets-management as key pitfalls.
WORDS WORTH SAVING
5 quotesWe started at 10:00 and finished about 3 o'clock in the morning.
— John Blackman
I just took and copied what Claude had put together and put it into Replit… and then it started going bloo-whoop, and there it was.
— John Blackman
Sometimes he goes off on a rabbit trail, and I have to bring him back.
— John Blackman
It always doesn't work in production what works in development, and that was very frustrating.
— John Blackman
It's just like AutoCAD… A lot of my friends didn't want to learn AutoCAD… I was still having fun.
— John Blackman
QUESTIONS ANSWERED IN THIS EPISODE
5 questionsCan you share the exact initial prompt + Word-doc structure you gave Claude that produced the roadmap and user stories (what sections mattered most)?
John Blackman, a 91-year-old church volunteer, describes turning a manual paper-based “impact weekend” registration process into a full web app by using Claude to generate a roadmap, user stories, and requirements, then feeding that into Replit Agents to implement the system.
How did you decide the boundaries between “system admin” and “local church admin,” and what mistakes did you make designing role permissions early on?
The app supports multiple churches and events with role-based admin levels, participant registration via QR codes, service toggles, waivers and signature capture, and printable “impact passports” (PDFs) that attendees can bring to the event.
What did Replit choose for the database/auth stack under the hood, and what would you change now that you’ve seen how it behaves in production?
It also generates operational reports (demographics for pastoral follow-up, service usage counts, food pantry/lunch ordering, vision center details) and includes an oil-change workflow that looks up oil/filter specs from a VIN using an API.
For the VIN → oil/filter lookup: what is the precise flow and why use the OpenAI API there instead of a dedicated VIN decoder service (accuracy, cost, reliability)?
They discuss practical realities of agentic coding—fast progress and low cost (hundreds of dollars), but real deployment pain around dev vs production differences, PDF/email attachment issues, and secrets management—plus broader advice about embracing new technology as they did with AutoCAD decades earlier.
What was the root cause of the PDF attachment failing in production, and what exactly changed about the PDF format between dev and prod?
Chapter Breakdown
Meet John Blackman: 91-year-old “vibe coder” building for community impact
Claire Vo introduces John Blackman, a 91-year-old church volunteer who used Claude and Replit to build a full event-registration and management system. John frames the real-world problem: Impact Weekends with many free services and a paper-based registration workflow that needed to scale.
A career of learning: engineering, AutoCAD pioneer, and “growth mindset” arc
Claire shares John’s unusually broad background—electrical engineering, running a hardware store, airplane mechanic certification, AutoCAD adoption, fiber optic projects, and even helping with Google Fiber. The chapter sets up why John is comfortable learning new tools late in life and draws a parallel between AI today and AutoCAD adoption decades ago.
How the church app idea started: turning handwritten workflows into software
John explains that the project began as a desire to stop writing everything by hand for registrations and event logistics. He drafted an outline of what he wanted, then sent it to Claude with explicit instructions to target Replit as the build environment.
Using Claude to create a roadmap and PRD-style requirements
The team begins with a broad prompt—asking Claude how to create a development roadmap—then lets Claude ask clarifying questions. From the outline and Q&A, Claude produces a structured MVP plan, phases, and a requirements-style document that can be handed off to an agentic builder.
The “Impact Passport” concept: participant-facing artifact + QR-driven signups
John introduces the “Impact Passport,” inspired by another church’s paper process, as a printable guide for attendees that lists their chosen services/stations. The passport idea evolves into QR-code-based registration that makes it easy for participants to sign up from flyers and phones.
User stories and roles: designing for pastors, ministry leaders, and attendees
Claude generates user stories that John finds largely accurate, with minor edits as development progresses. The conversation highlights how the system must serve multiple stakeholders—participants, local church admins, system organizers, pastors—plus non-functional needs like usability and reporting.
Multi-tenant architecture: system admins vs. local church admins (RBAC)
The app’s structure becomes notably sophisticated: a multi-tenant, role-based system where system administrators can see all churches and approve admins, while local church admins only see their own events and data. Claire calls out the complexity—this is a real “enterprise-style” access model built from a community need.
From spec to build: handing Claude output to Replit Agent
John describes the handoff: copy Claude’s structured requirements into Replit, then let the agent generate the codebase rapidly. Claire notes the project’s “meaty” structure (docs, migrations, generated assets), signaling a full-stack application rather than a simple form.
Admin interface demo: managing churches, events, services, and participants
John tours the system admin login and shows how it drills into individual churches and events. Admins can view participants, toggle services on/off per event, and generate operational reports used both before the event and for pastoral follow-up after.
Operational reporting deep dive: food pantry, vision center, oil-change logistics
The system produces specialized reports tailored to different service stations. Food pantry/lunch reports help order supplies; vision center reports capture participant details; waiver flows reduce liability; and oil-change reports combine check-in tracking with parts planning based on vehicle info.
Participant registration flow: mobile-friendly forms, VIN lookup, waiver + signature
The participant journey starts with scanning a QR code, opening a mobile form, selecting services, and completing required fields. For oil changes, the app performs a lookup and then collects a waiver acknowledgement and a captured signature—ending with a completed registration.
Adding features after launch: volunteer tracking + name-tag generation
John explains how new needs emerge once the core system exists—like printing volunteer name tags on label sheets instead of using sticky notes and markers. The change requires extending the schema (new volunteer table/model) and adding new admin flows, which Replit Agent implements in large chunks.
Managing AI “rabbit trails”: prompting, checkpoints, and stopping the agent
John describes the practical reality of agentic building: sometimes the AI goes off-track, so he intervenes with simple commands (“wait,” “stop”) and redirects. A key technique is reverting to a previously working point in history (a checkpoint mindset) and reapplying known-good code.
From development to production: deployment friction, PDFs, email attachments, secrets
The hardest problems appear during production deployment: behaviors that work in development may fail in production. John recounts debugging email delivery with attached passport PDFs (solved by changing PDF format) and ongoing issues with environment variables/secret keys (OpenAI key caching) affecting VIN lookup in production.
No prior coding background: why “no preconceived notions” can be an advantage
John clarifies he hasn’t coded traditionally (no TypeScript), aside from some logic scripting in CAD contexts. The group observes that not being opinionated about stack choices can help: John accepts the agent’s defaults and focuses on outcomes, while more experienced engineers may overconstrain the solution.
Cost, timeline, and real-world impact: a complex system built fast
John reports surprisingly low costs—hundreds of dollars total, with early rapid progress in the first two days—and contrasts it with estimates of months of traditional development. Church leaders are impressed, eager to use it, and the system promises to reduce admin burden while enabling better follow-up ministry.
Lightning round: product feedback for Claude/Replit and advice on embracing AI
John’s “asks” include smoother production deployments, better secrets management, and clearer UX for finding Claude chat history. He closes with practical optimism: AI can be scary, but used correctly it helps people and expands what volunteers can build—echoing how learning AutoCAD extended his career and enjoyment.
EVERY SPOKEN WORD
Install uListen for AI-powered chat & search across the full episode — Get Full Transcript
Get more out of YouTube videos.
High quality summaries for YouTube videos. Accurate transcripts to search & find moments. Powered by ChatGPT & Claude AI.
Add to Chrome