How It Works

Devotionals are prepared weeks ahead of their date. For each date, software first assembles the calendar context: the season, the day of the week, Christian and public holidays, and which passages and themes appeared recently.

An AI model then chooses from an approved catalog of verse references. It returns only references, never scripture text. Software removes passages used recently, fragments that would distort their surrounding context, holiday mismatches, and themes that have appeared too often.

The application looks up the exact wording of the chosen passage in a stored public-domain Bible text. An AI model then writes the English devotional from that passage. Context notes, when present, draw only on related passages from the same open Bible texts (the World English Bible and the Reina-Valera 1909), not on third-party publications.

A separate model call, with different instructions, evaluates the draft independently for faithfulness to the passage, accuracy of any context note against the passages it draws on, neutrality between traditions, pastoral risk, originality, and tone. Automated checks also confirm that the scripture matches the stored text exactly and that no prohibited claim appears. Any failure rejects the draft; thresholds are never lowered to fill a date.

Each accepted English devotional is adapted into Spanish and Brazilian Portuguese by separate model calls. The scripture is never machine-translated: each language uses its own stored Bible text.

The adapted version is then checked against the English meaning, including names, numbers, and references, and scored for fluency, cultural fit, and faithfulness. A language that fails is held back without blocking the others.

If no devotional passes every check, the system falls back in this order:

• Retry within a strict limit.
• Choose a different eligible verse.
• Use a devotional from a reserve that already passed the same checks.
• Publish the scripture alone, with an honest notice that the reflection is unavailable.

Accepted devotionals are published on their date at a permanent address. The preview image shown when a devotional is shared is rendered from the stored, verified text and reference; it is not generated by an AI model. If a published devotional needs a material correction, a new revision is created and the change is recorded on the corrections page.

Reader comments are also moderated automatically before they appear. A comment is published only when every automated check passes, and nobody, including administrators, can bypass those checks to publish a comment that failed them.

The service uses models from OpenAI. For security, we do not publish prompts, internal configuration, or credentials. Each published devotional keeps a private record of how it was produced so problems can be traced and corrected.

These safeguards reduce risk; they do not make AI output infallible. A devotional can still contain an error, an awkward phrase, or an interpretation you disagree with. If you notice something, please tell us through the contact page.