v1 to v2: what changed in how your voice profile gets built

for the past month, hold your voice asked you to describe how you write.

onboarding was a flashcard quiz. you picked between apple and nike for "voice style." between slack and mckinsey for "tone." between brené brown and goldman sachs for how you build an argument. ten questions, ten brand-flavored multiple-choice answers. our ai turned them into ten voice keywords. that was your voice profile.

it worked. it just wasn't yours.

the gap i missed for too long

six weeks ago i was using my own product on a real draft. the rewrite came back sounding fine. competent. wrong.

i dug into the prompt. it was using the keywords i'd picked during onboarding. concise, direct, specific. what those words mean to me, what they sound like when i actually write, none of that was in there. the model had labels. it didn't have a voice.

here's the analogy that finally cracked it for me. you tell a chef "i like food that's bold and rustic and seasonal." they cook you something. it's bold and rustic and seasonal. and it's nothing like what you actually cook in your own kitchen.

what v2 reads instead

v2 is built around one switch. instead of asking what kind of writer you are, we read what you've actually written.

three samples is enough. paste a newsletter, drop an .md file, whatever you have. our ai reads them and pulls out the keywords with quotes attached, the sentence rhythm, the opening phrases you keep using, the words you never use, and three anchor passages. those anchors are the highest-signal item in the whole pipeline. every rewrite you ever ask for now sees them.

the anchors aren't content templates. they're rhythm reference. when the model returns "shipped the integration last quarter, engagement went up across the board," that rhythm came from your samples. not from a generic essay voice.

the regenerate button finally got memory

v1 had a regenerate button that knew nothing. you'd reject a rewrite, click again, and get a new attempt that didn't know what you just rejected. half the time it was the same vibe with different words.

v2 stashes every rejection. each new attempt sees the rejected ones with explicit instructions: do not repeat these patterns; diagnose what they likely disliked. first regenerate jumps to a stronger model. second regenerate adds a "take a different angle" instruction. and when you finally accept, we extract a one-line lesson from the diff between accepted and rejected, then feed that lesson into every future rewrite prompt.

after enough cycles, the rewriter stops fighting you about rules you don't care about and starts pushing harder on the ones you do. it's a slow loop. you feel it after a week, not a session.

gate b: the receipt

the part that took the longest to build is invisible until it isn't. every rewrite now gets scored on a 0 to 100 voice match against your anchors. below 70, we escalate to a stronger model with the failure reason. if both passes still don't clear 70, you see a "low confidence" badge in the diff tooltip. it's a small thing. it tells you "this is the best we got, scrutinize it" instead of silently shipping a bad rewrite into your draft.

and a bento

last thing. once you've written enough (ten documents or 2,500 words, whichever comes first), the analysis tab opens with a 9-card snapshot of how you write. your archetype. peak writing time. most-used word. your catchphrase. ai accept rate. a 10-draft score trend. you can download it as a jpeg and share it.

the bento isn't the point. the bento is what falls out when you have an actual voice profile to look at. v1 had nothing useful to show you because the profile was a guess. v2 has a profile worth screenshotting.

what to do about it

if you're already a user, nothing's automatic. v2 is opt-in per account today. flip it on, paste your three best samples in the brandvoice tab, hit re-analyze. takes 90 seconds. that's the upgrade.

if you're new, onboarding still asks a few flashcard questions. the difference is at the end you'll see a step asking for samples. you can skip it. i'd ask you not to.

v2 is closer to what we should have shipped first. there's a v3 in my head where the rewriter learns from your latest draft, not your onboarding samples. but v2 is the right direction. v1 was a guess about the kind of writer you wanted to be. v2 is a description of what you actually sound like at 9 p.m. on a tuesday.

write something. give it three samples. tell me what you think.