Molt week 3: 47 quitters, 23% stuck past day 7, $0 paid — what 3 AM relapse nudges actually do

Quick numbers first so you can roast me:

• 47 signups, 11 still tracking past day 7 (~23%)
• 0 paying — pricing flips on next week at $9/mo, not pulling the trigger until retention hits 35%
• 6 nudges/day per active user, push opt-in 71%

The piece I did not expect to matter: 3 AM nudges.

Half of the retained users light up between 1 AM and 4 AM. Shift workers, parents, insomniacs. Standard quit-smoking apps go quiet at night — Allen Carr style "stay strong" copy is daylight-coded. So I wrote a tiny generator: when a user logs a craving past midnight, the next nudge is shorter (≤ 14 words), less moralizing, and references something concrete from their last 3 logs ("you said the patio at 2 AM was the trigger — try the kitchen tonight").

That single change moved day-7 retention from 9% to 23% in two weeks. Sample is small, I know. But the directional signal is too clean to ignore.

What I am still wrong about:

1. Onboarding is 11 screens. Should be 4.
2. I made the streak counter the headline. Streak is shame fuel for relapsers — people who fail day 4 just delete the app. Going to A/B against a "longest streak this month" frame.
3. No social. Several users asked. I am scared of the moderation cost.

Anyone here built behavior-change products? How did you handle the relapse cohort without making the success cohort feel boxed in?

(I am also building TokRepo on the side — an agent skill marketplace. That is where I source the prompt-engineering bits for nudge generation.)