the editor had a gap.
it could mark a phrase. it could say the phrase sounded generic, too polished, too soft, too long, or too far away from the profile. but a user still had to trust the flag without seeing the evidence behind it.
that is not enough. if the product says a sentence does not sound like you, it should also say what it used to make that call.
what changed
profile-backed feedback adds a small evidence layer to the editor. not a new workflow. not another screen. the same highlighted text now explains whether the signal came from the profile, the samples, the selected format, a known ai pattern, or the user's own accept and dismiss history.
the important part is where this appears. the tooltip is still contextual. it shows up when the user is looking at the highlighted text, not as a random floating lecture on the page.
the profile now has a readiness score
a profile can exist and still be thin. five vague adjectives are not the same as five real writing samples. so the voice profile screen now shows a readiness score: how much usable evidence the editor has.
the score is not meant to be gamed. it is a practical warning. if feedback feels generic, the next step is probably not "run more checks." it is add samples, build anchors, and give the system more accept or dismiss signals.
accept and dismiss now matter faster
the editor already had accept and dismiss actions. now those choices are treated more seriously. accepting a fix stores that rule as something the user agrees with. dismissing a suggestion stores the opposite signal.
that matters because brand voice is not static. a profile built in onboarding is a starting point. the real profile gets sharper when the user reacts to real suggestions inside real drafts.
voice memory is the visible part of that loop
the voice profile screen now has two memory cards: your voice and not your voice.
"your voice" is where repeated accepted fixes show up. "not your voice" is where repeated dismissals show up. that distinction is useful because not every rule should get stricter. some rules should get quieter for a specific user.
this also makes the product easier to debug. if the editor keeps pushing a suggestion that the user hates, we can see whether the product is learning that rejection or just repeating itself.
why this matters
most writing tools explain from a generic rulebook. hold your voice should explain from the user's own evidence.
that is the bar for this release. every serious flag should answer a simple question: "why did you say that?" if the answer is weak, the product should show the profile is thin. if the answer is strong, the user should be able to see the source right in the editor.
less mystery. fewer generic corrections. more honest feedback.
that is the point.