A wish generator is only useful if you trust what's on the other end. Here is exactly how ours works, what it sees, and what it never keeps.
You type a few words - the occasion, who it is for, a little context. A language model takes that and returns a first line you can edit. We don't own the model. We tune the prompts, shape the tone, check the output, and hand it back to you. That is the whole mechanism.
The prompt is yours. The final message is yours too - we just help you start.
When you hit generate, we pass a short prompt to the model API: the occasion, who the message is for, the tone you picked, and - if you uploaded one - a compact description of the photo. Nothing else travels out.
Your prompt and the reply are not written to our database. They are not in server logs. They are not shared with anyone beyond the model provider for the length of the request.
Every response passes a pre-send check for slurs, abuse, leaked personal data, and anything that does not belong in a wish. If it fails the check, we generate again or show a hand-written fallback.
We don't train a model and we don't fine-tune one. The base model comes from the provider. We shape its behavior with prompts only.
We hold up our side of the trust - private inputs, no storage, outputs we check before they reach you. In return, a small ask: treat a generated line as a draft, and be upfront about AI when it matters.
Treat the result as a draft, not a finished voice. Edit before you send.
Be open about AI involvement in professional, public, or journalistic writing.
