Church Production Vault

Everything you know about your church's production system stops living in your head.

Take photos of your rack. Drop in a manual, a quote, your meeting notes. Tell an AI agent to ingest it, and get back a structured, cross-linked wiki that answers questions six months later, with citations, and with an honest distinction between what someone confirmed and what the AI guessed.

It's a folder of plain text files on your computer. Nothing to sign up for, nothing to subscribe to, no server, no company holding your stuff. Free, MIT licensed, from Overflow Creative.

---

Start here

You don't need to know how any of this works, and you don't need to touch a terminal. You need one app, whichever you already have:

• Claude Code, the desktop app
• ChatGPT, the desktop app (the Codex side of it)

Both work the same way for this: you point them at a folder on your computer and tell them what you want in plain English.

1. Make an empty folder. In Finder, wherever you keep things. Documents is fine. Right-click, New Folder, call it church-vault.

2. Open that folder in the app. Claude Code: Select folder, pick it. ChatGPT: open Codex, add it as a local project.

3. Paste this in and hit enter:

Set up a Church Production Vault for me in this folder, using the template at https://github.com/jakeh280/church-production-vault

Don't use git for this. Download and unpack https://github.com/jakeh280/church-production-vault/archive/refs/heads/main.tar.gz so the files land directly in this folder, not in a subfolder. Then read CLAUDE.md and run the first-run setup interview on me, one topic at a time.

That's the whole install. The agent pulls down the template, then interviews you for about ten minutes about your church, your room, your team, and how work actually gets approved. It fills in the config, deletes the example content, writes your first real pages, and logs day one. You just answer questions.

The "don't use git" line is deliberate: git isn't installed on a fresh Mac, and asking for it triggers a system popup you don't need. If your agent says it can't download at all, do it yourself in ten seconds: click the green Code button at the top of this page, Download ZIP, unzip it, drag what's inside into your folder, and then say set this up for me.

Then use it. Walk out to your booth and take ten photos: the front of the rack, the back of it, the patch panel, the labels, the console, whatever's confusing. Drop them in the inbox folder, and say:

ingest

That's the loop, and it's the fastest way to get a room documented, no typing required. The agent reads what's actually legible in the photos, writes the pages, and flags what it couldn't read clearly instead of guessing at a model number. Same loop for a PDF manual, a vendor quote, or your notes from a Tuesday meeting.

Worth doing, now or later: open the folder in Obsidian as a vault. Everything works without it, but Obsidian is what makes the links clickable, the callouts render, and the database views show up as sortable, filterable tables. Bases need Obsidian 1.9.10 or newer.

Later, when you want a backup: this lives only on your laptop, so a dead laptop takes it with it. Ask your agent to put it in a private GitHub repo and it'll walk you through it, secret guard and all. Private, always. Nothing about a production vault belongs in a public one.

---

Why this exists

Church production knowledge lives in three places: one person's head, a spreadsheet nobody updates, and a group chat from eight months ago. Mostly the head. When that person is out sick, the room doesn't run right. When they leave, it walks out with them, and the next person rebuilds it by trial and error on a Sunday morning.

That is the actual problem this solves. Not note-taking. Getting the system out of one person's head and into something the whole team can query.

The usual fix is "write documentation," and it usually fails, because documentation goes stale silently, and a wiki nobody trusts is worse than no wiki. So this template is built around the two questions that actually determine whether a page is worth acting on:

• Where did this come from? (source:): a person on a date, an ingested manual, a photo you took, or agent-inferred, which is the lowest trust level and is marked as such
• Did a human confirm it against reality, and when? (verified:), and only a human ever sets this. An AI can write a page; it cannot verify one

Blank is an honest answer. A page that says "nobody has checked this" is more useful than a page that quietly implies someone did.

The rest of the design follows from one rule: every fact lives in exactly one place, and every other page links to it. Two pages that both state the gain structure are two pages that will disagree the day one gets updated.

The bones of this are Andrej Karpathy's LLM wiki pattern: raw sources in, a compiled wiki out, three operations, ingest, query, and lint. This is that pattern pointed at one specific room, with the parts a production booth actually needs, gear lifecycle, spend history, verification, and a hard line around what a church must never file in a folder like this.

---

What it does

Ingest: you drop source material in inbox/, photos, manuals, quotes, meeting notes. The agent reads it, tells you what it found before writing anything, updates the pages it affects (rather than creating a fifth page about the same console), updates the index, logs the action, and archives or deletes the original depending on what's in it.

Answer: you ask a question. It answers from your pages with citations, not from general knowledge about church audio. When the vault doesn't know, it says so and creates the stub. Every unanswered question becomes the backlog for what to document next.

여기까지가 README 앞부분입니다. 나머지는 원문에서 읽을 수 있습니다. README 전문 보기 ↗