AI Sorry Card Generator

Three steps, about thirty seconds.

Open the card builder and press Write it with AI. The panel appears under the message box.

Pick who the card is for. Twelve options cover partners, friends, family, and work, and the choice changes both the wording register and the design.

Describe what happened. This is the part that matters. Type the specific thing you did, with a day attached if you have one.

Press generate. The message appears in the card, the design updates to match, and everything stays editable.

The model has no facts about your life. Everything it writes comes from the sentence you typed.

Type I was mean to her and you get something that fits any argument anyone has ever had. It reads as competent and hollow, which is worse than clumsy and sincere.

Type I cancelled on her twice in three weeks and never explained why and you get a message naming a real event, because you gave it a real event.

The pattern holds across every generation. Vague input, vague card. Specific input, usable card.

This is why the box asks what happened rather than offering a dropdown of generic scenarios. The specificity has to come from you either way. AI shortens the writing, not the thinking.

Six constructions are blocked at generation time, and they are the same six every guide on this site warns about.

It cannot use the word but, because everything before it gets cancelled by what follows.

It cannot start a clause with if, so you will never get sorry if you were upset.

It cannot use passive voice like mistakes were made, which removes the subject from the sentence.

It cannot apologize for their reaction, so sorry you feel that way is off the table.

It cannot ask for forgiveness, because an apology gives something rather than requesting something.

It cannot describe how badly you feel, since that shifts the focus onto you.

The design is not decorative guesswork. It follows one rule: scale decoration down as the situation gets more serious.

Small everyday mistakes carry warm colour and a sticker comfortably. Being late, forgetting something minor, a bad mood that landed on someone.

Broken trust, betrayal, or repeated behaviour gets a plain card with no sticker and no heart. Heavy decoration on real hurt reads as an attempt to charm your way past it.

The heart only ever appears for a romantic partner. A heart on a card to a manager or a friend introduces a note that distracts from the message, so it is forced off for everyone else regardless of what the model suggests.

Work, manager, client, and teacher cards always come back plain and pale. That is enforced on the server rather than left to the model, because a cheerful apology to a client is a real mistake and not a stylistic preference.

Read it before you send it. That sentence is the whole of the advice.

Change anything untrue. The model works from one description, so it fills gaps with plausible detail that might not match what happened.

Change anything that does not sound like you. The person receiving the card knows your voice, and text outside your normal range is the exact signal an apology cannot survive.

The design stays editable too. Every theme, colour, shape, and toggle works exactly as it did before, so a design you disagree with takes one tap to change.

If the draft is close but not right, generate again. Repeated runs produce genuinely different wording and design rather than the same card twice.

Building a card manually involves no network request at all. Names, message, colours, and export all happen on your device.

The AI writer is the one exception, and it only runs when you open the panel and press generate.

What gets sent is the situation you typed and the relationship you picked. Your name, their name, and the finished card are never transmitted.

The description goes to NVIDIA, which runs the model. Nothing is stored on our side, and no account exists to attach it to.

If that tradeoff does not suit the situation, write the message yourself. The message library has twelve human-written examples covering partners, friends, family, and work.

Four situations where writing it yourself produces a better card.

When the apology is serious. Broken trust deserves your words, and the effort of writing them is part of what the apology proves.

When you have apologized for this before. A repeat apology needs you to name the pattern, and naming a pattern honestly is not something you delegate.

When the relationship has a specific voice. Long friendships and marriages have a register that generated text misses.

When you already know what to say. If the sentence is in your head, type it. The AI exists for the blank box, not for the sentence you already have.

Generation is capped at eight requests per ten minutes from one connection.

That is generous for someone drafting a card a few times and tight enough that scripted abuse hits the wall quickly. Running a free AI endpoint means the quota is shared, and a cap is what keeps it working for everyone.

If you hit the limit, the manual path still works. Presets, themes, colours, and the message library are all unaffected, because none of them touch the network.

Output is capped at 180 characters, matching the card. The model cannot pad the message with filler even if it tries.

A template gives you a structure with someone else's facts in it. An AI draft gives you a structure with your facts in it.

Both need editing. The difference is what you are editing: a template needs the situation swapped in, a draft needs the details corrected.

Templates win when your situation is common and the example is close. The twelve in the message library cover most everyday apologies well.

AI wins when your situation is specific and no example fits. That is most serious apologies, because the specific detail is what makes them specific.

Neither wins if you send it unedited. A template sent as written reads worse than a plain text, because the design implies effort the words do not support.

The honest answer depends on what you do with the draft.

An apology proves two things: that you understand what you did, and that you thought about it. The understanding comes from your description, which the model cannot invent. The thinking is what happens when you read the draft and decide whether it is true.

Sending a draft unread fails both tests, and the person receiving it will feel something is off even if they cannot name it.

Using the draft as a starting point and correcting it does not fail either test. You supplied the facts and you made the judgement calls. The model handled sentence construction, which is the least important part.

The uncomfortable version of this question is whether you would tell them you used AI. If the answer is no, edit the draft until the answer changes.

Llama 3.1 8B Instruct, running on NVIDIA NIM.

It is a small model by current standards, which suits this task. The card message is under 180 characters and follows a fixed four-part structure, so raw reasoning capability matters less than instruction following.

The constraints do most of the work. The banned constructions, the character cap, and the design rules are enforced in the prompt and again on the server, so the output stays inside the shape regardless of what the model produces.

A larger model would write slightly smoother sentences. It would not know any more about your situation, which is the actual limiting factor.

The card marks the moment. The change does the repair, and that holds whether you wrote it or edited a draft.

Do the thing you named. An apology followed by no change becomes a data point, and the next one gets read against it.

Do not follow up asking whether they saw it. That converts a gift into a request for a reply.

If they respond with more detail about the harm, that is the best outcome even though it feels worse. Listen and do not defend.

If they say nothing, wait. Silence is not rejection, and a second message into it reads as pressure.

The AI writer is free with no account, like everything else on the site.

That is possible because the card itself is generated in your browser, so the only server cost is the AI call. Rate limiting keeps that bounded.

There is no paid tier holding back better output. Everyone gets the same model, the same constraints, and the same eight generations per ten minutes.

No watermark appears on anything, whether you used the AI or wrote it yourself.