Until it Passes

What it does

You type what is going on, in your own words. It tells you the feeling makes sense, then gives you one thing to try. Not a list. One. There are two separate spaces. A feeling I can't steady covers anger, panic, going numb, being cruel to yourself, sadness that will not shift, 3am worry: fifteen cards. An urge I don't want to follow covers porn, scrolling, the late-night pull and the hours after: thirteen. Twenty-eight in all, plus two step-by-step flows with almost no words on screen, for when reading is too much. Every card says what it is for, what to do, and where the technique came from. Nothing you write leaves your phone, and there is no account to make.

Why it works this way

One rule shapes everything: if you have to pick from a menu, read a wall of text, or make a decision, it has already failed. The reason is simple. The moment this is built for is the moment you cannot think straight. When you are that wound up you cannot hold much in your head, and that is exactly when every other app hands you forty options and asks you to choose. Privacy works the same way, and is not a legal box to tick. The things people type here are the things they would never say if they thought it was being stored somewhere: porn, hating yourself, not wanting to be here. If it left the phone, people would write a softer version — and a softer version is one the app cannot match a card to.

Under the hood

Three checks run before anything else: someone in crisis, someone talking about drugs or alcohol, someone describing a medical emergency. They are plain code rather than AI, they run on your phone, and they work with no signal at all. The same three exist twice, once in the app and once on the server, tested against the same list of phrases, with a test that fails if the two ever stop matching. They are set to fire too easily on purpose. Stopping someone who was not in crisis just shows them a phone number and a way back. Missing someone who was does not get a second try.

What the testing caught

260 tests in JavaScript and 149 in Python, run against the same list of real phrasings: 49 for crisis, 17 off-topic, 8 medical, and 23 that must not set anything off. Then eight reviewers each took one part of the app, and everything they found was handed to a second reviewer whose job was to prove it wrong. Seventy-nine findings went in, sixty-five survived. What they found is the useful part. The crisis check missed the most ordinary ways people say it: "i wanna die", "i wish i was dead", "i took all my pills" all went straight through. Worse, a sentence that only hinted at a stuck thought matched a keyword confidently enough to serve a card instantly, so the crisis check never saw it. And the urge section promised nothing you say leaves this phone directly above a box that sent it to a server. All fixed. But the pattern is the real finding: none of it was broken code. Every one was a place where what the app promised had got ahead of what it actually did. That is the worst kind of bug in something asking a person at their lowest to trust it.

Not a replacement

It is not therapy and is not trying to be. All 28 skills were checked against the research. Seven have decent evidence behind them, seventeen have thin evidence, three have very little. That sounds bad. It is really just the honest state of this research, and saying so out loud is the point. Here is why. DBT works and there is real evidence for it, but as a whole package: weekly therapy, a skills group, phone support, over months, with an actual clinician. Almost none of it has been tested the way this does it, which is one skill, on your own, on a phone, in ninety seconds. Worry postponement is the clearest example. Proper trials exist, and every one tests doing it daily for a week or more, not once at 11pm on a bad night. So every card tells you where its technique came from. Of 118 sources, 90 were checked and confirmed. The other 28 are labelled unconfirmed rather than quietly passed off as fact.

Still missing

No clinician has read the cards yet. They have been checked against the research and picked apart by reviewers, which is not the same thing. Two need a real one most: the card for the middle of an episode, whose main idea has never actually been measured by anyone, and the newest card, on jealousy. There have also been two testers. Every phrase the app matches on was made up by me or by a machine, and it has already been caught mishandling real sentences that no made-up list produced. The crisis check is a keyword list, not a screening tool, and does not pretend otherwise. It does not yet cover the second or third time in one night, the flatness of early sobriety, being scared of hurting someone else rather than yourself, or being mid-argument. What it does give you meanwhile is the skills, and a record that is yours: any range of dates, out as a plain text file, to read back or to hand to a therapist.