Product · 4 min read
Working With RecentReborn on Voice That Sounds Like You
The hardest problem in this product is not delivery. Sending a message when somebody comments is the easy half, and every tool in this category does it.
The hard half is that the message has to sound like the person it is sent from, to an audience who has read hundreds of things that person actually wrote. That is where automated replies get caught, and it is what we have been working on with RecentReborn.
What the problem actually is
A model asked to write a friendly reply will write a friendly reply. It will be grammatical, warm, slightly over-punctuated, and completely unlike how anybody talks to a regular customer.
The tells are consistent:
- Register drift. A formality nobody uses in a DM. "We would be delighted to assist you with that" from an account whose owner writes "haan bhej deti hun".
- Uniform length. Every reply the same three sentences, because that is what "write a helpful response" produces. Real writing is lumpy.
- Manufactured warmth. "So lovely to hear from you!" reads as personal once and as a machine on the twentieth send.
- Answering a category. A paragraph about the brand when somebody asked whether it comes in navy.
None of those are fixed by a better model. They are fixed by having a record of how a specific person writes, and by refusing to send when the record does not cover the situation.
What a voice profile is
It is built from replies the account owner actually wrote. Not a tone slider, not a personality setting, not a prompt describing a brand's values. A record of real sentences, used to shape generated ones.
That design has an honest cost, and it is worth stating plainly: a new account with nothing written has nothing to build from. The first drafts read generic — exactly the thing the feature exists to avoid — and fixing it means pasting in a body of your own answers, which is work you have to do.
Where the collaboration focused
Three things, all of them about restraint rather than fluency.
Knowing when not to generate. The most valuable behaviour is declining. A model that is unsure will still produce a confident sentence, because confidence is what it produces. Detecting that and stopping is harder and more useful than improving the sentence.
Grounding. A reply should be traceable to something you gave it — a price list, a policy, a knowledge source you wrote — rather than assembled from plausibility. An ungrounded answer that happens to be right is still a system you cannot trust twice.
The queue rather than the send. When confidence is low, the reply waits for a person instead of going out. That trade is invisible when it works and obvious when it is missing.
What we did not build
No voice cloning in the sense people sometimes fear — nothing here imitates a person who has not chosen to be imitated, and the only input is the account owner's own writing, which they can delete.
No claim that generated replies are indistinguishable from human ones. Often they are not, and the accounts that do best with this are the ones that template the repeated questions — free and unlimited — and reserve generation for the genuinely unusual message.
What it changed in the product
The visible outcome is a shorter, more conservative generated reply that defers more often. Some of that reads as the model being less capable. It is the opposite: the confident version was easy and the restrained one required knowing where the edges are.
The gate's content safety check is the last of ten, and it is worth knowing that it judges our output, not your customer's message. It exists to stop us saying something on your behalf that you would not have said.
Where this goes
The direction is more refusal rather than more fluency: better detection of the messages that need a person, better grounding on the ones that do not, and a clearer record afterwards of which was which.
If you want the full anatomy of a voice profile, that has its own post. If you want the argument for templates over generation, it is here.
