Privacy and AI

What happens when you ask for a draft

A plain-language trace from your browser to generation, fallback, editing, rehearsal, and export.

Implementation reviewed 16 July 2026

Before you press generate

Use initials, placeholders, or broad descriptions. The writer intentionally leaves personal fields blank and offers a clearly synthetic sample. Full names, exact locations, health details, secrets, and details about other people are not needed for a useful structure.

Generation flow

Only fields for the active writer are included in the request. Vow requests do not send speech fields, and speech requests do not send vow fields.

The browser sends a bounded application/json POST request with a versioned schema.
The server validates content type, byte size, allowed tone and length, text limits, and rate policy before a provider call.
If configured, Vercel AI Gateway routes the request to the selected model provider. User text is delimited as data and instructed not to override drafting rules.
Provider output is schema-validated. If the provider is absent or fails, the server uses a built-in template and labels it accurately.
The response is private and no-store. Editing and version history then happen on your device.

What AI can and cannot know

The system is instructed to use only facts you supply and not to invent memories, qualities, dates, places, cultural rules, or religious requirements. Automated safeguards can still fail. Treat every generated sentence as a suggestion and remove anything you cannot personally confirm.

Religious and cultural wording on this site is general writing guidance, not officiant, legal, theological, or cultural advice. Verify ceremonial language with an appropriate human adviser.

Abuse and cost controls

Generation and export use layered IP, local-session, and instance-wide checks inside each warm serverless instance. Identifying network values are hashed in memory and are not written to application logs. Browser session identifiers can rotate or reset, and separate serverless instances do not share these in-memory counters, so the checks are best effort and do not create a durable per-person, per-IP, or project-wide quota.

There is no durable overall cost ceiling configured in this release. Distributed enforcement through a platform firewall, gateway budget, or shared rate-limit store is a precise deferred P1 infrastructure boundary; until that exists, the site must not describe the current checks as guaranteed global cost control.

Return to the productOpen the private writing workspace

Generation flow

Only fields for the active writer are included in the request. Vow requests do not send speech fields, and speech requests do not send vow fields.

The browser sends a bounded application/json POST request with a versioned schema.

The server validates content type, byte size, allowed tone and length, text limits, and rate policy before a provider call.

If configured, Vercel AI Gateway routes the request to the selected model provider. User text is delimited as data and instructed not to override drafting rules.

Provider output is schema-validated. If the provider is absent or fails, the server uses a built-in template and labels it accurately.

The response is private and no-store. Editing and version history then happen on your device.

What AI can and cannot know

Religious and cultural wording on this site is general writing guidance, not officiant, legal, theological, or cultural advice. Verify ceremonial language with an appropriate human adviser.

Abuse and cost controls