You've seen an AI invent a Bible verse. In five short lessons you'll catch it doing that with your own code — and then fix it, yourself, so it can't.
By the end, the AI on your computer answers Scripture questions only from a real Bible it looks up. And you'll open a real production project and recognize every part of it — because you just built the small version with your own hands.
You finished concord-tutorial-web (or you can fetch from Concord and put the result on a page), and that's all. No AI experience needed — not the apps, not the APIs, none of it. If you've heard of ChatGPT and you're not sure what's actually inside, you're exactly who this course is for.
Already a developer? Skim freely — this is a beginner's course, deliberately unhurried. If you want the production-grade version of what it builds, go straight to concord-mcp.
Found this because Scripture matters to you? Maybe you've personally watched a chatbot quote a verse that doesn't exist — chapter, number, and all. That's not a reason to walk away from these tools. It's the exact problem this course teaches you to fix: by the last lesson, every quote arrives over the wire from a real Bible on your own computer, and you'll have written the code that makes it so.
This is the third Concord course. Course 1, concord-tutorial-web, taught you that a server is something you can ask questions. Course 2, concord-tutorial-react, taught you that real code is something you can read — it's recommended, not required; this course is plain HTML and JavaScript throughout.
Two 30-second checks, both straight from your browser's address bar:
http://localhost:8000/healthz— a short line of data means Concord is on (course 1's ritual, unchanged).http://localhost:11434— the words "Ollama is running" mean the AI is on.
If either check fails — or you've never installed Ollama — that's what SETUP.md is for. One sitting, and every lesson after just works.
- Another local server — talk to an AI from a page you control, then ask it for John 3:16 word for word — and discover you can't tell right from blurred without checking. ← start here
- Catch it making things up — five claimed verses, five live Concord lookups, your own marking tally. Ours came out 1 of 5.
- Give it a tool — the heart of the course: a ~40-line loop and a one-sentence rule (without it, our model looked things up in 2 of 6 runs; with it, 6 of 6) — and your tally becomes 5 of 5 for a reason you can read.
- Let it choose — three tools, and the model routes by reading descriptions you wrote (ten for ten in our test). Then you sabotage one and learn what production already knows.
- It has a name — a chat that remembers, worth showing someone — then concord-mcp, read at its tag, until you recognize your own hands in production code.
When you've finished all five, the curious can keep going: concord-mcp is the production implementation of exactly what you built — its README shows the no-code path to running the real thing. Graduates also get recipes.md and ideas.md — steal freely.
MIT © 2026 Kris Bennett — see LICENSE. (Parity with Concord.)
Never trust a quote without a wire line.