Inspiration

I've spent ten years in customer success leadership watching new managers walk into the hardest moments of their careers with zero reps. You wouldn't give a big presentation without a run-through, but we hand people a termination conversation and wish them luck. The talk they dread happens in the shower, at 2am, in the car. I wanted the practice to happen somewhere it counts.

What it does

Rally lets a manager practice a hard conversation out loud before having it for real. You pick the conversation you've been putting off ("Feedback that isn't landing," "Demote your work best friend," "Deliver a raise you don't agree with"), and an AI plays the other person live. They go quiet, talk over you, deflect, cry. When the scene ends, a critic reviews your exact words: what landed, what got lost in hedging, whether you held your pauses. It counts every "just," "sorry," and "maybe." You can run it again until the real one goes better. A 60-second Conflict Style quiz ("The Over-Apologizer," "The Slow Boil") tells you which conversations to start with.

How we built it

Expo / React Native, built almost entirely in pair-programming sessions with Claude Code. Claude Sonnet plays the bosses (fast, in-character); Claude Opus writes the reviews (specific, kind-not-nice). ElevenLabs gives each boss a distinct voice; speech-to-text happens on-device. RevenueCat handles the paywall: a Rally Pro subscription, plus a $4.99 "Big Conversation Tomorrow" pass for the night before it counts. API keys live behind a small relay server on Railway so nothing sensitive ships in the binary.

Challenges we ran into

The biggest one wasn't technical: halfway through, the app's original theater framing ("Tonight's Cast," battle cards, curtain up) started to feel wrong. These are real, painful conversations, and dressing them as a show cheapened them. We tore out the whole metaphor a week before feature-complete and reframed everything around the conversation you have coming. Technically: making the pause meter work (measuring the silence you hold after a hard sentence is a genuinely weird engineering problem), and keeping AI bosses difficult without making them impossible.

Accomplishments that we're proud of

The reviews are actually good. The critic quotes your exact words back to you, and testers feel seen ("You apologized four times. She heard zero of your feedback."). In our eval pass, we fed it deliberately vague manager-speak and it caught the missing message on all twelve bosses, each with a different specific receipt. Every scenario is authored by someone who has sat in the real conversation.

What we learned

Clear is kind. In app copy, in feedback design, in the product itself. Every time we made the app more honest and less clever, it got better.

What's next for Rally

Reverse Mode: the AI plays the bad manager, and you practice as the employee. Asking for the raise. Pushing back on scope. Quitting gracefully. Everyone with a boss has a conversation they're putting off, too.

Built With

Share this project:

Updates

Submission history