In the Weeds: Can You Write Your Wedding Vows With AI?

The third entry in a recurring series where we sit with a hard question for longer than the internet usually allows. The first entry was about parents and homework — what happens when a parent opens a chatbot at 9:17 pm on a Tuesday to help a kid get through a worksheet. The second was about students and learning — what happens at 2 am to a graduate student staring at a sentence on page nine that has, for the fifth time tonight, refused to mean anything. This one is about a couple. More precisely, about one half of a couple. More precisely than that, about the half of a couple who got the job of writing the vows, which is to say the half who is currently sitting in a small pool of lamplight at 11 pm the night before the rehearsal with a chat window open in one tab and a blank document open in another and a feeling in their stomach they have not felt since the night before a college exam they hadn't studied for.

It's 11:04 pm. The rehearsal is at 4:30 the next afternoon. The wedding is the day after that. Somewhere in the apartment — or the hotel room, or your sister's spare bedroom — your partner is asleep. They were asleep early because they had a hard week. You told them, three weeks ago, that the vows were "almost done." They were not almost done. They are not, currently, started. There is a Google Doc called vows_draft_FINAL_v3.docx that contains exactly three things: your partner's full legal name, the phrase "from the moment I saw you," and a single comma you do not remember typing.

You have been trying to write the rest of the document for forty-five minutes. You have written and deleted the second sentence eleven times. You have, in this exact order: gotten up to make tea, sat back down, opened a notes file from 2019 to see if past-you wrote anything you could steal, closed the notes file, opened a wedding-vows article on a magazine website that wanted to load 47 trackers, closed the magazine website, opened a different tab.

The new tab has a blinking cursor in it, in a chat window, in whichever AI has slowly become the one you reach for when you don't know what to type. The tea is going cold.

The question underneath the question is the one this piece is about.

The question underneath is: am I about to cheat on the most personal sentence I will ever say out loud, or am I about to use a tool the way grown-ups use tools?

We at a-gnt have been sitting with this one for a while. The first In the Weeds was about a kitchen table. The second was about 2 am. This one is about something quieter and, in a strange way, harder than either. The stakes are not a grade or a worksheet. The stakes are a sentence you are going to say in front of every person who has ever loved you, to one person you intend to love every day for the rest of your life, and the question of where the sentence came from is going to be a thing you carry whether anyone else knows or not.

We are not going to give you the easy answer. The easy answer comes in two flavors and both of them are wrong.

The first easy answer is no, obviously not, vows are sacred. People who say this are right about the part that matters and wrong about the rest. They're right that there is a thing in the vow that has to be yours — has to come out of your specific life with this specific person — and they're right that an AI cannot manufacture that thing. They're wrong about how writing actually works. The vow doesn't appear, fully formed, from your soul. It is written. Writing is a craft. People who write for a living use editors, beta readers, prompts pinned to the wall above the desk, and every other scaffolding tool humans have invented for getting unstuck. The "vows are sacred so don't touch a tool" argument confuses the content of the vow with the process of arriving at it, and the confusion causes people to write worse vows.

The second easy answer is yes, why not, it's just words. People who say this are thinking about the wrong wedding. The wrong wedding is one where the vows are a formality everyone forgets by the time the salads come out. There are weddings like that, and if you're at one, fine — type "write me five romantic vows about my fiancée Sarah, she is a nurse, we like hiking" into a chat box and have a great life. Most weddings, though — and most of the people we talk to who are stuck on this — are not that wedding. They are having the wedding where the vows are the part of the day that is going to live in their head and their partner's head for years. For that wedding, "it's just words" is the wrong frame, because those particular words are what the relationship is going to remember.

So neither easy answer survives. The actual situation is harder. Here is the thesis we want to earn:

An AI can help you find what you mean. It cannot write what you mean. The two operations look identical from the outside and they are not the same operation, and the difference is the whole game.

That's the sentence we're going to be circling. We're going to walk through a hypothetical composite — a person at 11 pm with a chat window open — and look at what the AI does well, what it does badly, and a two-pass framework that doesn't lose the part that has to stay yours. By the end you will know what to type, what not to type, and where the AI has to leave the room.

A note before we go in. The composite is a composite — not one couple, not a quote, just the shape of a thing we have watched happen with enough variations that we can describe it honestly without inventing biographies. When we say "you," we mean the person reading this piece who is in the situation. You are the composite. The composite is you.

What the AI does well, when it's used like a tool and not like a writer

Here is the part nobody who is mad at AI writing wants to admit: the early innings of the composite go better with the chat window open than without it.

Not because the AI is going to write the vow. Because the part of vow-writing that comes before any sentence appears on the page is the part the AI is structurally good at. And almost everyone who is stuck on their vows is stuck in that part — the part before the writing — without knowing it. They think they are stuck on the writing. They are stuck on the finding, and the writing is just where the stuckness shows up.

There are three things, specifically, that a chat window is unreasonably useful for at 11 pm the night before a rehearsal, and we want to name them so you can recognize them when you're doing them.

The right interview questions. A blank Google Doc is a terrible interviewer. A blank Google Doc asks one question — "what do you want to say?" — and then sits there judging you for not having an answer. A chat window, prompted right, is a different kind of room. You can ask it to interview you. Not "interview me about my fiancée." That is the wrong prompt and you will get back a paragraph of corporate-feeling questions. The right prompt is something like "please ask me a series of small, specific, oddly angled questions about my partner — questions a thoughtful friend would ask if they were trying to help me remember a moment I'd forgotten — and ask them one at a time, and wait for me to answer each one before moving on. Don't compliment my answers. Don't summarize them back to me. Just ask the next question."

What you get back is a different experience entirely. The model asks: what is something your partner does when they think no one is watching? You answer. It asks: what was the smallest argument you've ever had — small enough that it shouldn't have mattered, big enough that you remember it? You answer. It asks: what have you learned to do because of them, that your past self wouldn't recognize? You answer. What's a sound in your apartment that didn't exist before they moved in? You answer. What's a word your partner says wrong, that you've started saying wrong on purpose? You answer.

After twelve or fifteen of these, you have something a blank Google Doc was never going to give you: a small heap of specifics that are unmistakably about this person. None of them came from the model. The model only asked. The answers came from inside your head, where they had been all along, buried under three weeks of reception logistics and the wrong kind of pressure. The model is not the writer here. The model is the friend across the kitchen table at midnight who knows how to ask the question that surfaces the thing you forgot you knew. It is exactly the kind of work 🔓Unstick Your Creative Brief was built for in other contexts, and it translates cleanly here.

Cutting the dramatic bloat. Second thing the AI is good at, and it surprised us. People at 11 pm writing wedding vows do not, mostly, write too little. They write too much. The first draft, written by a person who is nervous and trying hard, almost always overshoots: it gets dramatic, then cinematic, then tries to use the word forever in a way that is both unearned and slightly wet. It is a wedding-vow-shaped object made out of vow material from movies. It is the vow your panic wrote.

A chat window, used with the right instruction, can cut this in a way you cannot do alone at midnight. The instruction is: "here is a draft of my wedding vows. Please flag any sentence that sounds like it could appear in any wedding vow about any person. Do not rewrite. Do not improve. Just point at the sentences that are generic, and then stop talking."

What the AI gives you back is a list of the bloat. "Sentence 2 is generic. Sentence 5 is a phrase you've heard at four other weddings. Sentence 9 contains the word 'soulmate,' which does not appear anywhere else in your draft and is doing no work." You read the list. You wince. You knew it was right while you were writing it, and you only kept it because you thought a vow was supposed to sound like that. It isn't. A vow is supposed to sound like you, and the AI, used in this narrow way, is a faster and less embarrassing way to find the parts that don't than reading the draft aloud to a person you love.

This is what good editors do. They don't write the better sentence. They cross out the sentence that isn't yours and leave the page open for you to come back and write a real one. The chat window can do exactly this, at 11:34 pm on a Wednesday when no editor on earth is awake.

Surfacing the right shape for a feeling. Third thing, the most subtle, the most worth being careful about, because it's the closest to the line. Sometimes you have a feeling about your partner you cannot find a sentence for — not because it's too big, but because it's specific in a way that resists any of the obvious shapes. You don't want to say "I love you because you make me laugh," because that sentence is going to be in seven other vows at the same venue this week. You want the actual thing — something like the way your partner specifically makes you laugh — the small private noise they make when something dumb happens to a stranger — has, over four years, become the sound your nervous system reaches for when something has gone wrong. That's the sentence. You didn't have it ten minutes ago. You had a feeling.

A chat window can help you find a sentence for a feeling — if you ask the right way and if you do not let it write the sentence for you. The right way: describe the feeling in three or four messy sentences. Tell the model you are not asking for the sentence; you are asking for help naming what the feeling is about. Ask it for three different angles on what you might be trying to say, and explicitly ask that none of them be a finished sentence — no quote marks, no draft language, just angles. You read the angles. One is closer than the others. You think about why. The thinking about why is where the actual sentence shows up, in your own voice, because you wrote it with a brain that has been freshly nudged.

The risk here is enormous. If you ask for the sentence, the model will give you the sentence, and it will be smooth, and the smoothness is the trap. Smooth is the enemy. You are trying to write something with the small irregular shape of your particular love for this particular person, and small irregular shapes do not survive a sentence-smoother. Angles only, never finished sentences. Write me the sentence turns the model into the writer, you into the editor of the writer, and the vow into a paragraph from a movie. Your partner will feel it the moment you start reading.

What the AI does badly, almost every time, for free, without being asked

Now the other side. We named three things the AI does well. We are going to name three things the AI does badly, in a vow context, and we are not going to soften them, because the failure modes are the part that get people hurt.

It writes sentences that sound like everyone's vows. The most consistent failure mode, and it's structural — the same way it was structural in the Hallucinations entry on originality. A model trained on the internet's corpus of wedding-adjacent text will produce, when asked for vows, the central tendency of all wedding vows. The central tendency is the part of all wedding vows that is most like all the other wedding vows. That is exactly the part you do not want. You can fight the geometry with prompting, with examples, with rewrites — but you are always fighting it. The model's path of least resistance is the centroid, and the centroid is the place a vow most needs to not be.

You can spot centroid output without trying. It uses the word journey. It uses partner in crime. It contains a sentence that begins from the moment. It contains a sentence ending for the rest of my life. It mentions adventures. It mentions home. None of these things are wrong individually. Together, they are the dead giveaway. A real vow has at least one sentence the centroid would never produce — about the way your partner pronounces "lasagna," about a fight in an IKEA, about what their dad said the second time you visited their house.

It misses the specific weight that only you can know. Every long-term relationship has at least one thing in it that means more than a stranger could understand from outside. A phrase your partner once said in a hospital hallway. A specific Tuesday in November. The way they showed up to a thing they did not want to show up to, for reasons you and they don't need to spell out. The vow, if it's going to do its job, has to land on at least one of these — not necessarily by naming it, but by being aware of it, shaped around it, a sentence that you and your partner both know is about that one moment even if no one else does.

A model cannot do this. It cannot do this even a little. It does not know about the hospital hallway. If you tell it about the hallway, it will produce a sentence about a hospital hallway in the way a stock photo is about a hospital hallway. It will not be about the hallway. The weight of the actual thing lives between the two of you, and the only person who can put it into a sentence that carries the weight is you. This is the part where the AI has to leave the room. Not negotiable. Not something a better prompt fixes. We have looked.

It can't feel the relationship. A wedding vow is not a piece of writing in the usual sense. It is a piece of writing produced by a body that has lived inside a particular relationship, in a particular set of weeks and months and years. The body remembers things the head does not. Vows are spoken by a body, in a room, to another body, in front of bodies that love both bodies. A model has none of this. A model has tokens. The tokens can imitate the form of a vow, but the form is not the thing. The thing is what happens in your throat when you are saying the sentence and suddenly realize you mean it more than you knew you meant it ten seconds ago. The model cannot produce that, because the producing and the meaning are happening in different places, and the place where the meaning happens is the place the vow has to come from.

The two-pass framework

So now you know what the AI is good at and what the AI is bad at. The remaining question is how to combine those facts into a way of working that gets you the benefit of the tool without losing the part that has to stay yours. The framework is two passes. We are going to be specific, because if we are vague the framework collapses.

Pass one: the AI is the interviewer. You are the only writer in the room.

Open the chat. Use this exact instruction, or one shaped like it:

“I am writing my wedding vows. I am going to read them aloud at my wedding in two days. I do not want you to write any vows. I do not want you to draft any sentences. I do not want you to suggest any phrases I could use. Your only job, in this entire conversation, is to ask me one small specific question at a time about my partner and our relationship, and to wait for my answer before asking the next one. Do not summarize. Do not compliment. Do not reflect my answers back. Just ask the next question. Ask me twelve to fifteen questions, total, and then stop. The questions should be small, specific, slightly oddly angled, and should reach for the kinds of memories I would not have thought to write about on my own. Begin.

You answer the questions in a notes file, not in the chat — this is important. The chat is the interviewer. The notes file is where the answers live, in your voice, with whatever digressions the answers want to take. Do not edit while you are answering. Just answer the questions, twelve to fifteen of them, in the messiest, most-specific way you can. When the AI runs out of questions, close the chat. Walk away for ten minutes. Make actual tea and let it steep.

When you come back, the chat is closed. Open the notes file. Open your draft. Now you are the only person in the room. Read your answers. Two or three will stand out — a memory you hadn't thought about in months, a moment you wrote down in the rush that, on the second read, makes you stop. Those two or three are the vow. The rest is scaffolding. Build the vow out of them. Write it the way you would write a letter to your partner if you knew the letter would be the only thing they had of you. Write it badly first. Cross out the bad parts. Read it aloud in a whisper. If you wince at a sentence, cut it. If you cry at a sentence, leave it. Crying is not a reason to cut. Wincing is.

This is the part where the AI is not in the room, and it is the part where the vow gets written. The model asked the questions. The questions surfaced the answers. The answers surfaced the vow. None of the words in it are the model's. That's the line.

Pass two: the AI is the de-bloater. You are still the only writer in the room.

When the draft is roughly there — when you have a version you are not embarrassed by and not happy with, which is the right state for a draft to be in — open the chat one more time. New conversation. Use this instruction:

“Here is a draft of my wedding vows. I am not asking you to rewrite. I am not asking you to improve. I am not asking you to suggest alternatives. I am asking you to do one thing: read each sentence and tell me whether it sounds like it could appear in any wedding vow about any other person. If it could, flag it as generic. If it has a specific detail that anchors it to this relationship, leave it alone. Do not write anything else.

Paste the draft. Read the flags. They will sting. Go back to the draft, with the chat closed, and rewrite the flagged sentences. Not with the AI's help. With your own — by going back to the notes file and finding a more specific memory the generic sentence was reaching for. The sentence wanted to say you make me feel safe; the more specific version is the one about the night the apartment lost power and your partner found the candles in the third drawer without having to ask. Write the second version. It's harder and slower and it's the version the vow deserves.

Then close the chat for the last time. The vow is done. Read it aloud one more time, alone, in the room. If a sentence makes your throat tighten in a way that has nothing to do with embarrassment, it stays. If a sentence makes you want to skip past it, it goes. That's the only test that matters.

A few tools, gently

The tool that does most of this work is mostly a chat window with the two prompts above. Three things in the a-gnt catalog came up for a reason and are worth naming.

📖The Draft Reader is a soul we built to be the patient, honest reader you don't have at midnight. It's exactly the right register for the second pass — it will tell you what is generic without trying to fix it.

🔓Unstick Your Creative Brief is the prompt we built for the moment when the writing has stalled and you can't tell why. The diagnostic move — a few questions, a name for the knot, one small experiment — is the same move that helps when the vow has stopped moving.

🎙️Creative Voice Coach we mention with an asterisk. The skill will not generate content in your voice. It will read what you wrote and help you see it. That is the right register. The wrong register — imitating your voice, writing a sentence you're tempted to use — is exactly the failure mode this piece has been about, and the Voice Coach was built to refuse it. Use it for the seeing, not the writing.

11:52, the same room, the same lamplight

It's 11:52 pm now. The tea is gone. The chat window is closed. The notes file has fourteen messy answers in it, two of which you are going to use. The draft on the other tab is no longer empty. It contains six sentences. Three are bad. Two are okay. One is the sentence that, when you read it aloud just now in a whisper, made you stop and look up at the ceiling for a few seconds, because the sentence is true in a way you had not let yourself say out loud before tonight.

That sentence is the one. Everything else in the vow is going to be built around it. It came from inside you, surfaced by a question the model asked, written down by your own hand, in the room you are in. The model did not write it. The model is, right now, on the other side of a closed tab, and it does not know about the sentence, and it never will, and that is correct.

In the morning you'll read the vow once and not touch it. You'll read it again at the rehearsal and almost mean it. You'll read it again at the wedding itself and mean it so much you'll be surprised at how much you mean it. Your partner will hear it and look at you in a way you'll think about for a long time afterward. None of the people in the room will know what tools you used to write it, and that is also correct, because the tools are not the point. The point is that you wrote it.

The first In the Weeds asked: am I cheating for my kid right now, or am I helping? The second: am I becoming a worse student right now, or a different kind of student? This third one asks something quieter, because the audience is just you and the person you are about to marry.

The question is: did I write this, or did the machine?

The answer is: you wrote it. The machine asked the question that helped you remember what you meant. The remembering and the meaning and the writing all happened in your body, in your voice. The machine left before any sentence appeared. That is the line. As long as the line holds, the vow is yours. As soon as the line breaks — as soon as any sentence in the vow originated outside your head, even a little, even with edits — the vow is something else, and your partner is going to feel it, and you are going to know.

Hold the line. Use the tool for the finding. Do the writing yourself. Close the tab before the writing starts. Read the vow aloud in the lamplight, alone, one last time. Then go to bed.

We'll be in the weeds again next month.