Second Take

Every turn has three replies: one that works, one that backfires, and one that sounds fine but has a hidden cost. The first two are easy to write. The third is where the work is.

"Whatever you think is fair, honestly. I trust you" sounds generous in a raise conversation. It also hands the number to someone who will anchor low to stay safe on budget. "It'll be fine, Papa. Don't worry so much" sounds kind, and it tells a worried father you haven't thought about the risk. Obvious wrong answers teach nothing; the believable ones are the lessons.

Choices are shuffled every turn, so the good answer isn't always in the same place.

Each partner has a hidden want, a list of things that make them soften, and a list of things that make them dig in. Marcus, the manager, wants to keep you but needs evidence he can take to HR. Papa is scared, because he watched his nephew's startup fail.

Under the hood, mood runs from -3 to +3 and falls into one of three states: guarded, neutral or open. Every partner line is written three times, once for each mood, so the character stays the same person but warms up or closes off. Paths merge back together after each turn, which keeps the writing manageable: about three versions of each line instead of a tree that doubles every turn.

On desktop, a live panel shows the partner's triggers and highlights the one your last reply hit. A real app would probably hide this until the end. I left it visible because it shows the thing I actually designed: not just what the character says, but why.

Every choice carries a one-sentence coach note that names what the line did, and a better line to try whenever the choice wasn't the strong one. Each scene has a rubric of three or four skills (for the raise: backed it with impact, named a clear ask, asked without apologising, left with a next step), and every choice is tagged against it.

At the end you get a score, a breakdown by skill, the two skills to work on next, one line to rehearse, and a sentence about your pattern, such as "You tend to skip acknowledging the other person's pressure before stating your need." The rule I held to was no generic praise: every note had to say what the line actually did.

Each scene is one JSON file, separate from the code, so rewriting a line never means touching the app. A validator runs on every build and checks the writing's structure: exactly three choices per turn, one of each quality, a better line for every weak choice, every mood written for every line, and every trigger label matching the partner's lists. Structural mistakes in the writing fail the build, the way a spelling mistake would fail a spellcheck.

XP, streaks and sessions are stored in the browser. On desktop the app sits in a phone frame with the live panel beside it; on a phone it runs full screen. I built it in Cursor.