A small piece of software that takes anonymous text from people you know and turns it into a portrait of you is — let's be honest — a slightly weird thing to exist. We think about this a lot. Here's where we've landed.
The four things we think hard about
1. The friends, not the owner
Every other "feedback" product in this space treats the friend as a means to an end. They are a data source that produces a number for the customer. We don't think of it that way. The friends are the people whose words are being interpreted by a model and turned into a paragraph that someone they love is going to read in a coffee shop on a Tuesday.
So we made structural commitments:
- The owner never sees who answered. Not which devices, not which IPs, not which order, nothing.
- We rate-limit submissions per IP per mirror so a single person can't write five different "voices."
- We never use the answers to train any model — not ours, not anyone's.
- We encrypt every answer at rest with a key that doesn't appear anywhere a database dump would expose.
2. The model is interpreting humans, not measuring them
We asked the model to do one specific thing: synthesize what was actually said. Not infer. Not psychoanalyze. Not score against a framework. The system prompt explicitly tells it to refuse to invent traits the answers don't support.
This is partly an ethics call and partly a product call. The reports are better when the model stays close to the words. They are also more honest. We don't want to ship a product that confidently tells you something about yourself that no one actually said.
3. The owner shouldn't be hurt by their own report
The contradiction question, the blind-spot section — these are the parts where the report can sting. We tested versions that pulled punches. They felt useless. We tested versions that didn't pull punches. They felt cruel. The current version sits somewhere in the middle: it will name a real pattern, but it will name it the way a thoughtful older sibling would. Direct, never mean.
We've been wrong about that calibration before. If a report ever lands wrong, we want to hear it — every report has a feedback link.
4. The whole thing should be deletable
You can wipe a mirror, your answers, and the report, on demand, by emailing us. There are no accounts to chase down, no "soft delete" — the encrypted ciphertext is hard-deleted from the database.
What we won't do
A short list of things we have explicitly decided not to build, even though they would make the business easier:
- No "send my report to my manager" or "share with my therapist" feature. Those involve other people in a thing they didn't agree to.
- No leaderboards, no scores, no "your friends rate you 7.2 out of 10."
- No selling anonymized data. There is no anonymized version of "what your friends said about you" that should ever leave this database.
- No selling ad inventory against the report content. The report is a private object. It should stay one.
The thing we are most proud of
The most common feedback we get is some version of: "I cried, then I texted them." People read what their friends actually wrote and then they reach out and say something they've been meaning to say. That is what we want this product to do. If it ever stops doing that — if it becomes a personality test, a corporate 360, a content engine — we will have lost the plot.
If you want the receipts: our privacy page documents the technical details, and the example report shows what the output actually looks like.