How I Built an AI Team for Church Communications
I clicked on a YouTube video because I thought the title was wrong.
Note-Taking Apps Are Dead. That felt like bait. I’ve been living in Obsidian for years — before that, Bear, before that, Byword. I’ve been writing in Markdown longer than most people know what Markdown is. The idea that note-taking apps were dead seemed like a hot take for clicks.
I was right that the title was wrong. I was wrong about what the video was actually going to say.
I should back up.
I’m the Communications Director at New Hope Church in Brandon, Florida. My job is to take ministry vision and turn it into messaging that reaches people — campaigns, social media, the website, sermon content, weekly emails. One person. A lot of channels.
For a while, my solution was ChatGPT. It helped. I’d built up a solid library of context documents — the church’s theology, mission statement, voice guidelines, staff structure — and fed all of it into projects so the AI had something to work with. The writing was better than starting from scratch.
But the workflow was a grind. Ask a question, get an answer, copy it out, paste it into Obsidian, spend five minutes cleaning up the formatting because something always broke in transit. Every campaign. Every week.
The other problem was context. Every time I started a new campaign, I was starting from nothing. I’d have to manually load old files, remind the AI what we’d done before, explain how this year’s Easter connected to last year’s Easter. The AI was a good writer. It just had no memory. Custom GPTs helped speed it up, but not enough.
That YouTube video introduced me to a different way of thinking about it. Not note-taking apps are dead — that part was still wrong. But the idea of building a dedicated team of AI agents, each with their own role and their own voice, instead of asking one generic AI to do everything?
That landed.
I knew Claude existed. I’d looked at it. But I’d already put real time into my ChatGPT setup — the context documents, the projects, the whole system I’d built — and switching felt like starting over. The pain of staying has to outweigh the pain of changing before you actually move. Eventually it did.
So I ran a small test: I gave Claude access to my Obsidian vault and pointed it at a problem I’d been wrestling with for months.
We have a church-wide planning document — a live Word file that lists every event on the calendar. Staff update it constantly. The problem is, the moment you copy it anywhere, it’s out of date. I could never get a stable version into my vault without it immediately going stale. And because I couldn’t annotate it or link it to campaign notes, I kept re-reading it every week trying to remember: did I already deal with this event? Does this need marketing? Did I talk to someone about this?
I asked Claude to convert it to Markdown, search my vault, find every campaign note that already existed, and wire in wiki-links.
It did. And it found the right notes — not by exact filename, but by context. It figured out what I was asking for. That was the thing that made me sit up.
None of this is as foreign as it sounds.
You probably already automate more than you realize. You ask Siri for directions rather than unfolding a map. Your bills pay themselves. There’s an out-of-office reply handling your email while you’re gone — a thermostat that learned your schedule without you explaining it — a podcast app that downloaded the new episode before you even knew it was out.
None of that felt like a philosophical shift. It just made sense. AI is the same thing, just further up the stack.
So I built the team.
I named them after nerdy characters — Jarvis is the orchestrator, Vex handles social media, Samwise writes blog posts, Tuvok handles sermon content, and so on. Partly because it’s fun. Also because names like “Social Media Bot” would never stick in my head. My coworkers give me funny looks when I talk about it. My wife asked, sincerely, “What do you mean you have a staff?”
I didn’t have a clean answer. I still don’t, really. But here’s what I know it actually does.
The writing sounds like us. I’d already built those context documents — church theology, voice guidelines, mission and vision. Claude didn’t just have them as reference files. It used them to shape each agent’s voice from the inside. The social copy sounds like New Hope. The blog posts sound like me writing about New Hope. Before, the AI had the documents. Now, the team has internalized them.
The social media writer knows what else is happening. I’m notoriously bad at posting the same copy on Facebook and Instagram and then realizing the link doesn’t work on Instagram. Vex knows that. She also knows what’s on the calendar for the week — so when she’s deciding what to highlight, she’s looking at the whole picture, not just the campaign I handed her. She mixes in non-promotional content, faith-engagement posts, things worth sharing that have nothing to do with an event. That’s a real shift. I’m not using AI to write more promotional posts faster. I’m using it to write less promotional content, because the system knows enough to do it right.
Campaigns have memory. When we run Vacation Bible School this summer, the team has access to notes from every VBS before it — what worked, what the graphics looked like, what we promoted and when. I don’t brief it from scratch. It already knows.
Here’s what I showed a colleague a few weeks ago: a social media draft, before I’d touched it. She read the Instagram caption for Fifth Sunday Kids in Service — “We have no idea exactly what’s going to happen — and that’s exactly the point.” — and laughed. It worked as straight invite copy. It also worked as an inside joke for anyone who knows what Fifth Sunday at New Hope actually looks like. She was genuinely surprised the draft was that good before I’d edited it.
My workflow now: I read the output, make notes in the margin, dictate my edits back to Claude, and it rewrites. I don’t re-edit by hand. I correct the direction and it executes.
I still use Obsidian. The note-taking app is not dead. But I’m not spending my Sunday afternoons building Dataview queries and MetaBind fields to make my vault display data correctly. I just tell Jarvis what I need and it happens.
Couldn’t have told you that a year ago. Not the writing part — I already knew AI could do that. The part I didn’t see coming: that it would know everything I’d already done, understand what else was happening that week, and sound like my church when it wrote.
I clicked on a video because I thought the title was wrong.
It was. But it changed how I work anyway.
Stuart Mackey is the Communications Director at New Hope Church in Brandon, Florida. He writes about church communications, technology, and the weird overlap between the two.
