Personas
AI coaching personas built from public figures' words, served as a real-time chat platform.
The Problem
Most AI personas are knowledge-retrieval ("What would X say?"), not coaches that probe, challenge, and drive to action. Building authentic coaches from first principles requires systematic extraction, behavioral design, and grounding in real conversational dynamics — not vibes.
What I Built
Two halves of a system. (1) An 8-phase pipeline (Source ID → Acquisition → Extraction → Taxonomy → Module Synthesis → CLAUDE.md → Testing → Cleanup) that turns a public figure's body of work into a 4K–10K token behavioral blueprint plus 15–25 situational modules. (2) A Next.js 15 web app with real-time streaming chat (Vercel AI SDK + Claude Opus 4.6), invite-gated access, usage tracking, module loading tools, and Postgres persistence. Seven personas shipped: Jeff Bezos, Esther Perel, Paul Graham, Matt Mochary, Chris Camillo, Stanley Druckenmiller, Jerry Colonna.
Notable
The system's ground truth is an "11 Elements of Coaching" framework reverse-engineered from real coaching transcripts. Anti-patterns override LLM sycophancy. Example interactions show behavior over description. Pattern detection catches real-time avoidance. The framework earns directness through trust rather than asserting it.
Stack
Status
Open source at github.com/blaizew/personas-web.