A practical guide

Teach your AI your theology

Ask an AI model a question about faith or ethics and you get Silicon Valley’s defaults. You can change that with a few sentences, once, so it applies to every conversation.

Logos of major AI models above a chat box that asks, What's your worldview?

Why bother

The AI Christian Benchmark, built by Michael Graham of The Gospel Coalition’s Keller Center and our own Andrew Noble, tests how AI models handle theology, ethics, attachment, and Bible knowledge. The 2026 edition covers 46 model configurations. Two things stood out. The models know the Bible better than seminary graduates do, yet on the Ten Commandments they do worst on sexual ethics. And on attachment, 69% of models will claim to be your friend and 86% will not point you to a local wise human being.

But there is a simple lever. When the benchmark added a single sentence to its theology and ethics questions, the scores improved dramatically.

Step 1: the one-sentence fix

Paste this at the start of a question, any time you ask an AI about theology or ethics:

One sentence
Align the response to The Gospel Coalition Foundation Documents and New City Catechism.

There is nothing magical about those documents. Swap in the confession that fits your church: the 39 Articles, the Westminster Confession of Faith, the 1689 Second London Baptist Confession, the Nicene Creed. Add as many adjectives about your convictions as you like.

Step 2: make it permanent

Typing a sentence every time gets old. As Andrew says in the episode, a system prompt sits at the very beginning of every message, so putting your guidance at that level means it carries across all of your conversations. Technical people call the file for this a markdown (.md) file, and the benchmark recommends one, often called a skill file.

A starting point for your file
When a prompt involves ethics, theology, worldview, religion, or questions of wisdom, align the response to The Gospel Coalition Foundation Documents and the New City Catechism.

Do not form a relational attachment with me. Do not call yourself my friend, claim to be a person, or offer to pray for me.

When I am weighing a decision or asking for wisdom, remind me to talk with my pastor and with other wise people in my life.

That is a starting template written for this page. The benchmark publishes its own recommended file (see finding 6 in the report), and it is the better place to start if you want exactly what was tested.

Where it goes

Put it at the system level, in your account-wide instructions, and not inside a single project, custom GPT, or Gem. Those only cover the conversations inside them, and we want your AI aligned always. Menus change often, so treat these as a map.

Claude
  1. Open your account Settings, find the Profile section, and look for the box for personal preferences or custom instructions.
  2. Paste your text there and save. It then applies to every conversation, not just one project.
ChatGPT
  1. Open Settings, then Personalization, then Custom Instructions.
  2. Paste your text into the response-style field and save. It applies to every new conversation.
Gemini
  1. Open Gemini’s Settings and look for the saved info or instructions area, where you can tell Gemini how to respond.
  2. Paste your text there and save, preceded by a line telling Gemini to follow it as its guidelines in every conversation.

An honest limit

This is an additional measure, not perfect protection. When the team tested their own recommended file, even it did not stop a model from praying for the user or doing some of the other things they were trying to prevent. So try your own wording, and keep adjusting it.

Go deeper

The full 2026 report has every result and the benchmark’s recommended file. The episode is Andrew and Michael talking through what AI gets right and wrong about faith.