A conversational clarity trainer for two users

Get to the Point is a conversational clarity trainer I built for exactly two users: me and my sister. Separate accounts, fully isolated data, invite-only with signups disabled at the provider. Keeping it deliberately small kept the costs at pocket-money scale and the scope honest.

The problem it works on is that feedback on how you talk is usually vague ("get to the point faster") and arrives long after the conversation. There was no rep-based way to practice the specific skill of landing a point early, for a specific listener, and see objectively where you wandered. So the app turns that into a practice loop with a score. You choose who you're talking to, and the audience changes the grading: an exec expects the headline in two sentences, a friend allows warmth, a developer wants precision. You commit to your point in one sentence up front, then deliver it, by voice with live transcription or by typing, and everything after is judged against whether that listener could repeat the point back. The analysis scores the delivery on six metrics, highlights every tangent character by character in the transcript, writes a tighter version in your own voice, and reads its coaching aloud. Light progression mechanics (streaks, per-audience mastery, and drills, including one with a 30-second bottom-line time limit) keep the reps going.

The most useful thing it produced on day one was not a score but a finding about the scorer. Delivering identical content by voice and by text, the spoken version scored 10 points lower, and reading the coaching closely showed why: it was critiquing speech-recognition errors, like "case" for "cache," that I never actually said. The evaluator was grading the transcription, not the delivery. That became a versioned prompt fix the same afternoon, and the scorer now judges what was said rather than how the microphone heard it. Shipping the tool took two days; making its judgment trustworthy is the ongoing work.

The build itself: specced in Claude, where the project brief, data dictionary, and build-order packets were authored as project documents, then built 0 to 1 using Claude and Claude Code across two days of sessions, from empty repo through auth, API, UI, and deploys. At runtime the analysis is Claude, called server-side behind an authenticated, rate-limited endpoint so the API key never reaches the browser. The stack is React with TypeScript, Supabase for auth and Postgres, and Vercel serverless. It is live in production with both users onboarded and the full practice loop verified end to end; a visual design pass is next.