1. The interview
Everything starts with eight questions. Which role you are speaking in, your name, the couple's names, your relationship and how long you have known them, two or three specific memories, how funny you want it, anything to include or avoid, and a target length.
The memory questions are the ones that matter. They are deliberately prompted — "a funny moment you shared", "a time they showed who they really are" — because open-ended questions produce adjectives and prompted questions produce scenes. A speech is only as specific as the material it is given.
The role you select changes more than a label. Each of the fifteen roles carries its own etiquette guidance and its own list of prohibited topics, which are injected into the writing instructions. A best man speech and a step-parent speech are written under materially different rules.
2. The structure rules
Every reception speech is written to the same arc, because it is the arc that works: a hook opening, story one, story two, a pivot to the couple, a sincere close, and a toast invitation. Vows follow a different structure entirely — second person throughout, three to five concrete promises, and no toast, because there is no toast at a ceremony.
Length is mapped explicitly rather than left to chance: 350–450 words for two to three minutes, 550–750 for four to five, and 800–1,000 for six or more.
A specific list of clichés is banned outright, in every variation: dictionary definitions, "for those of you who don't know me", "marriage is like" similes, and generic quotations about love unless you asked for one. These are the phrases that make a speech sound like it came out of a machine, and they are prohibited at the instruction level rather than filtered afterwards.
3. The safety rules
Wedding speeches fail in predictable ways, so the same categories are excluded from every speech we write: ex-partners, bachelor and bachelorette party material, anything crude, anything about appearance or weight, jokes implying the marriage will fail, and humour that humiliates rather than teases.
Anything you list under "absolutely do not mention" is treated as absolute. It is not softened or worked around — it is omitted, and we do not allude to it.
We never invent facts about real people. Where a memory is thin, the speech is written to generalise gracefully rather than fabricate a date, a place, a quotation or a name you did not supply. This is the rule that most distinguishes the output from a general-purpose chatbot, which will happily invent a charming detail that never happened.
4. The models
We use OpenAI models, and we tell you which. Final speeches are written with GPT-5.6 Terra, the higher-capability model, because the three full versions are the product. Free previews, rewrites and the free tools use GPT-5.6 Luna, which is faster and cheaper, which is what lets the preview be free.
Every model call happens on our servers. Your stories are sent to OpenAI to write the speech and for nothing else, and the API key is never exposed to your browser.
Both model choices are configurable, so if a better model ships we can move to it without changing anything about how the product works.
5. Where you are still required
We are explicit about this: a generated speech is a draft in your voice, not a finished performance. Three things are genuinely yours to do.
- Check every fact. We only use what you gave us, but you are the one who knows whether the detail is right.
- Read it out loud, at least three times. Almost every problem in a speech is audible on the first read and invisible on the page.
- Change anything that is not how you talk. Manual editing is unlimited and free precisely because this step matters.
We also publish a plain-language AI disclosure, because we think anyone reading a wedding speech site deserves to know exactly what they are buying.