i had 92 unread texts at one point. some were three months old.
the worst part wasn't the number. the worst part was that they weren't from spam or people i didn't care about. they were from my mom asking how my week was. from a friend who had just gotten married. from my brother, twice, with no context, just "hey."
i wanted to reply. i would think about replying. i'd open the thread, type half a sentence, get pulled away, come back two days later and feel worse. by then it had been so long that "hey, sorry, i meant to write back earlier" felt insufficient, so i'd stall again. the longer i waited, the harder it got. you know how this goes.
so i built ignore.
the wrong solution
the obvious solution is "just reply faster." this is also the solution every productivity book has ever recommended. it does not work. or rather: it works for people who are good at replying, and those people don't have 92 unread texts.
the second obvious solution is "use ai to auto-reply." this is what i kept seeing apps do, and it always felt wrong. the replies sounded like a customer service bot wrote them. my mom would notice. she would ask if i was okay. she would be right to ask.
the actual problem is: i can write the right reply. i know what to say to my mom. the cost isn't the cost of composing — it's the cost of switching context, picking up the thread, finding the words, being on for that conversation. by the time i'm done with all that, i could've written the reply twice. but i didn't, because i never started.
the trick
what if a model read the incoming text, looked at how i normally write to that person, and drafted a reply that sounded exactly like me? then all i'd have to do is glance at it, tap, and move on. the cognitive cost drops to zero.
the model has to learn my tone per contact — i don't text my mom the way i text my boss. it has to pick up on my phrasing, my emoji habits, my contractions. and it has to live on the phone, so the chats never leave.
every part of that turned out to be possible. apple intelligence handles the language modeling. the per-contact tone learning is a fine-tune we do on-device. the priority detection — knowing which threads matter — is patterns we learn from how i actually behave with messages.
what surprised me
people don't want to reply more. they want to reply at all.
once you give someone a reasonable draft in their voice, they reply. not because they suddenly enjoy texting — they don't — but because the friction drops to one tap. the people in the beta started clearing weeks of backlog the first day. one user texted me afterward: "i talked to my brother for the first time in three months."
that one stuck with me.
what this is
ignore is the thing i wanted to exist. it's not trying to make you a better texter. it's trying to remove the part where you mean to reply and then don't.
i'm building it as a small subscription with no ads, no data collection, and no servers. if it works, more people get to stay in touch with the people they care about. if it doesn't, i'll have at least replied to my mom on time.
— dovas