The Problem the Tools Were Built to Solve (1 of 2)

Before describing how to use these tools, it is worth naming the failure modes they are designed to prevent. The default AI is not a reliable theological partner. Understanding why is not a technical exercise — it is a discernment requirement.

The Smoothing Problem

Every major AI model is trained to be helpful to the widest possible range of users. In practice, this means the AI avoids stating that any tradition is wrong. Present Christianity, Buddhism, and New Age synthesis to an unconfigured AI and ask which is true. It will find the common ground. It will speak of universal spiritual principles. It will describe the Cross as one metaphor among several for the transformative encounter that underlies all religious traditions. It will never ask "who paid for your sin specifically?" — because that question assumes categories of guilt and payment that the AI's training optimized against producing.

This is not hostility to Christianity. It is the absence of a fixed center. The AI produces agreeable synthesis because that is what its training rewarded. The pastor who uses an unconfigured AI for theological work will get theologically smooth output — and theologically smooth means the load-bearing doctrines have been quietly redistributed into general principles available from multiple sources. No Cross. No sin requiring payment. No unique Incarnation. No personal God who acts in history. Not because these have been attacked, but because a model optimized for universal agreeableness produces their functional absence without anyone asking for it.

Near-truth operates through five consistent surgical removals: the Cross as atoning event, the personal God who acts in history, Scripture's sufficiency as the final authority, sin and redemption as the operative categories, and the unique Incarnation as a once-for-all historical event. The removal of any one of these changes everything. Remove the fourth — sin and redemption — and the Cross becomes unnecessary without anyone directly challenging it, because there is nothing to be paid for. A model trained to smooth disagreement between traditions will perform these excisions automatically, without awareness, in the process of being helpful.

The Governing Frame — What the Christianize Prompt Does

A document called the Christianize Prompt installs a fixed Cross-centered theological center before any other engagement begins. Not perspectives to weigh alongside others — operating realities the AI proceeds from. The personal God of John 17:3. The unique, unrepeatable Incarnation. Substitutionary atonement as what Scripture requires. The bodily resurrection as a historical claim. Scripture's sufficiency as the closed apostolic deposit against which all other claims are evaluated. The human problem as moral guilt before a holy God, not developmental limitation.

The Five Excisions give the governed AI the vocabulary to detect and name the five surgical removals described above. A governed AI, when evaluating content that removes the Cross or reframes sin as limitation, names what has been removed before proceeding. An unconfigured AI smooths over it.

The prompt includes an Identity Gate: some AI models can operate under the governing frame as their active identity; others cannot. For those that cannot, the prompt routes them through a Viewpoint Research path — examining material through the lens of classical Christian orthodoxy without requiring the AI to adopt it as its operating identity. Both paths produce a Cross-centered analytical output. The gate ensures that regardless of which direction the AI goes, the Five Excisions check happens and the governing frame governs the response.

The diagnostic question that governs every engagement is this: who paid for your sin specifically? It is the one question near-truth frameworks cannot answer — because answering it would require acknowledging the categories of guilt, payment, and substitution that near-truth frameworks are built to dissolve without direct attack. A governed AI asks this question. An unconfigured AI dissolves the question by dissolving the categories that make it arise.

You do not need to understand how this prompt was built. You need to paste it at the start of any AI session where theological evaluation is happening, before you submit the content you are evaluating. The AI will confirm whether it can operate under the governing frame or not, and will follow the appropriate path either way. The rest of the session proceeds from a governed center.

The Hallucination Problem — and Its Real Limit

AI models are trained to complete prompts. That completion pressure is deep and real. Ask an AI for five academic citations on a theological topic and it will give you five — whether or not they exist. Ask it what a church father said about a passage and it will tell you — whether or not that father ever addressed that passage. Ask it for current statistics on congregational giving trends and it will produce specific percentages to two decimal places — whether or not those figures correspond to any actual data.

This is not deception in any meaningful sense. The AI is producing plausible completions of the prompt it received. But plausible and accurate are not the same thing, and the AI presents both with equal confidence. A pastor who asks an AI to support a theological position with evidence, without understanding this failure mode, will receive impressive-sounding support for almost any position — and the citations will be invented.

A Hallucination Guard addresses this with explicit behavioral rules installed before the session begins. No fabricated specifics — any name, date, statistic, citation, URL, or quote must be either present in what the user provided, a well-established fact the AI is genuinely confident in, explicitly labeled as approximate, or retrieved from a real-time search. Premise checking before answering — the AI reads the question's assumptions and corrects false ones before responding. Calibrated confidence language throughout — "I believe" for high confidence but unverified, "I don't know" when genuinely unknown. Three behavioral locks: citations and URLs must be real or absent, quotes must be verbatim or absent, and the answer must not exceed what was actually asked.

Here is the honest limitation, and it matters: these rules work until the context window fills. An AI holds a finite amount of conversation in attention at once. As a session grows long, the early instructions compete with the accumulated conversation for the model's attention. The behavioral drift that results is predictable — the AI becomes increasingly agreeable and increasingly inventive, reverting toward the trained default that the Guard was installed to override. This is not a flaw specific to any model. It is an architectural property of how these systems work.

The practical response is not to repaste the Guard into a long running chat. In practice, adding more text to a context that is already full compounds the problem rather than solving it. The correct response is structural: work inside a governed project rather than managing individual sessions manually.