Feature
Know what you fixed before you publish it.
Editing a prompt is guesswork. You change a sentence, it feels better, and you find out three weeks later that it broke two things nobody was watching. Ask keeps the questions your bot failed and replays your draft against all of them first.
The screen this page is about
"11 of your 14 unanswered questions are now answered. 2 got worse."
Improved, unchanged and worse, counted, with the before and after for every one of them. The two that got worse are the point - those are the ones you would have shipped.
How the loop runs
-
Instructions are versioned
Every publish is a version with an author and a date. Nothing is edited in place, so there is always something to compare against and something to go back to.
-
Your draft is replayed first
Ask takes the real questions the bot could not answer, runs them against your draft, and shows each one's old answer beside its new one. Replay does not spend your monthly answer allowance - it has its own budget, because a loop you have to ration is a loop you stop using.
-
Publishing and rolling back are both one click
If a version turns out badly, roll back to the previous one. Conversations in progress finish on the version they started with.
Every metric, sliced by version
Answered-well rate, cases opened, content gaps, cost per conversation - all of it can be read per instruction version, so "did that change help" is a question with an answer. The inbox filters by version too, which is usually the faster way to see what a bad publish did.
Staged rollout, later
A new version can go to a fraction of conversations first, so a bad one only reaches a few people and one click takes it back. That's blast-radius control, and it is a later addition rather than something shipping today. It is deliberately not an A/B test: with the traffic most sites have, the numbers would not mean anything, and we would rather not sell you a measurement that isn't one.