backbone: Conversation Coach

The debrief: five scores, your own lines quoted back, and a rewrite of the weakest one.
The rehearsal. Each goal turns green when you actually land it.
Pick who you're up against. Two personalities are free and three come with pro.
Rehearse your own situation. What you type is saved only on your phone.
After the first debrief, a OneSignal in-app message asks about reminders.
Plans and trials come from the RevenueCat offering, in the buyer's own currency.
First run: four panes, one idea each.
The app icon, 1024 by 1024.

Most new managers have their hardest conversations live, on the first try, with the person across the table. backbone is where the first tries go.

You pick one of twelve conversations new managers dread. The analyst whose weekly report has been late three weeks running, after you quietly covered for it twice. The director who promised your team to a VP without asking you. The peer who presented your team's analysis with their own name on the deck. The raise you have to turn down. The performance conversation you have already put off twice. Each one tells you who you are, who is across the table, what happened and what you need to walk out with. Then the other person starts talking. With pro you can also type in your own situation, the conversation you actually have coming, and backbone plays that person instead.

They are played by Claude Haiku 4.5 and they stay in character. Before your first line you choose how they play it: cooperative, defensive, passive-aggressive, stonewalling or emotional. They answer what you actually said, and anything vague gets used against you. They also have somewhere to be. If you are still circling the point when their time runs out, they end the meeting and leave. A real conversation ends when the other person decides it is over, and we wanted practice to work the same way.

Above the chat sit three goals for that conversation. For the late reports they are name the pattern, give a concrete example and get a commitment. Each one turns green when you actually do it. When you drift, a coach card says in one line what is missing and offers a sentence you could say next. One tap puts it in the box, and you can edit it before it goes.

After three of your turns the debrief unlocks. Claude Sonnet 5 reads the transcript and scores you from zero to ten on clarity, specificity, empathy, firmness and outcome. It is told not to inflate, so four to six is the normal range and nine is rare. It quotes two of your own lines back with a sentence on why each one worked or failed, rewrites the weakest thing you said, and ends on a blunt bottom line.

Then you run it again, harder. The same person comes back defensive, and with pro the rerun keeps climbing, through passive-aggressive to stonewalling. Your scores stay on your phone, and after four rehearsals the home screen tells you whether you are trending up or slipping. If you allow notifications, backbone brings you back on an evening before a workday to rehearse once more before the real one.

Three rehearsals are free, with the full debrief and no account. The three free scenarios are one each from feedback, saying no and boundaries, and two of the five personalities are free. Pro opens all twelve scenarios, your own situation, the three harder personalities and unlimited rehearsals. Nothing in the app asks for a name or an email. Your messages go to our server and on to Anthropic to write the replies, and we do not keep conversations in a database. The situation you type in is saved only on your phone, and the app opts out of Android's cloud backup so it stays there.

backbone is an Expo React Native app for Android, live on Google Play since 23 September 2026 and on version 0.2.0 since 24 September. RevenueCat runs the subscription, with one entitlement called pro and an annual and a monthly plan from a single Google Play subscription. The paywall reads every trial length and price from the RevenueCat offering, so it matches Play in the buyer's own currency, and when you run out of free rehearsals or reach for something pro, a line under its title says which. OneSignal runs the reminder prompt and the pushes. A small Node server behind a Cloudflare tunnel holds the model key, limits how many turns one device can use in a day, and asks Anthropic for replies in a fixed JSON schema.

The worst bug we shipped was in the debrief. The model sometimes closed its JSON one brace early, and our parser, which read from the first brace to the last, threw the whole review away. Structured outputs fixed it at the source, and a tolerant reader and one retry now sit behind that.

It went live on 23 September. That day it had 11 installs, 18 customers in RevenueCat, no trials and no revenue. We would rather show you that than round it up.