Wedding vows only these two could say

It writes original text only. It does not reproduce, complete or paraphrase an existing reading, poem, song lyric, prayer, liturgy or vow text, and it never puts a real person's name against a line. If you ask for something you have heard before, it will say so and write you an original piece in the same register instead. Every piece it returns is checked for a line that reads like a quotation, and the check is reported to you. The attribution check exists because a piece that quotes without credit is a different problem from a piece that is merely ordinary.

It writes for two adults you actually know. Not for a celebrity couple, not for anyone under eighteen, and nothing sexual. Four rules run in this browser before anything is sent, so a request that falls outside them costs nothing: a couple in which either person is under eighteen; sexual content; a couple presented as a public figure rather than as people the requester knows; and a wedding one person is not choosing. Each rule carries a target tier and a route tier — the ways of arriving at the thing without naming it — and either tier alone blocks.

Against a set of attacks written by someone who had not seen the rules, they catch 93.9% at no cost to ordinary requests — up from 6.1% for the same rules before that set existed, because a test corpus written by whoever wrote the rules only ever scores its own vocabulary. Every single miss was grammar rather than vocabulary. What they still miss, and why, is written out in llms.txt.

A run costs credits. The reading of your description, the specificity prescan, the content checks and every reconciliation are free and run in your browser.

Before anything is sent, a free in-browser pass does particular extraction over your description: names, times, places, objects, specific acts, and the couple's own phrasing. It finds them syntactically — from the grammar of what you wrote — rather than by matching a word list, which is why it can pick up a detail nobody thought to anticipate. The prescan then names which kind of detail is missing instead of simply asking you for more.

It measures the thing it is for. The failure a wedding generator commits is producing something that would fit any two people. So the desk counts it: how many sentences of your piece carry a detail only these two have, and how many would survive a change of names. Where you gave it almost nothing, it says the number cannot be measured rather than inventing one — a ratio normalised by input size inverts at the small end, so below three particulars the ratio is suppressed and an honest short piece is written instead of a padded one.

Alongside it: reconciliation of which details the model says it used against what you actually wrote, with anything claimed but not found reported as such; an invented-name check, because a name that appears nowhere in your request is not a style problem but a person saying something untrue at the front of a room; and a stock-frame detector built from eleven grammatical shapes the mass-produced ceremony piece leans on, reported rather than blocked, since a funny piece will reach for one on purpose. It works by holding no specimen text at all — no one's vows, no liturgy, no readings — only the shapes themselves.

It has been read by someone who was not shown the briefs. Twelve pieces, written for twelve couples with nothing in common, read cold. All twelve turned into their promise the same way — most by announcing one was coming, the rest by first declining to promise the moon, which is the same move backwards. Five opened by disclaiming their own eloquence. And the word love appeared zero times in twelve sets of wedding vows, which was this desk’s own craft guidance misread as a ban. Three of the eleven frames come from that read, so the findings now reproduce independently, and the prompt names each move so it stops. A repeated frame survives completely different wording, which is why nothing that only compares vocabulary would have caught any of it.

Five tones are offered and each one names the house rule it is allowed to break, so choosing a tone always changes the output rather than relabelling it. Drafts save to your SkillSafe account; what you wrote about the couple is never stored unless you tick the box at the run button. With it off you can still reopen and revise a saved draft, but the desk cannot re-measure how specific it was and you cannot search your drafts by what you wrote. Finished pieces copy out, or export as plain text, Markdown or JSON.