Personas
Open-source pipeline that turns a public figure's own words into an AI coaching persona — a system prompt that coaches, not a bot that quotes.
The Problem
Most AI personas are knowledge-retrieval ("What would X say?"), not coaches that probe, challenge, and drive to action. Building an authentic coach from a real person's body of work takes systematic extraction and behavioral design — not a one-line "act like X" prompt.
What I Built
A seven-phase Claude Code pipeline — Source ID → Acquisition → Extraction → Taxonomy → Module Synthesis → CLAUDE.md Construction → Testing — that turns a public figure's writing, interviews, and talks into a character-first CLAUDE.md system prompt plus 15–25 situational modules the persona loads on demand. Parallel Opus agents handle extraction and synthesis; raw source material stays gitignored while a per-persona catalog records where each source came from. Each persona is a self-contained directory you run directly in Claude Code. Ten figures shipped: Jeff Bezos, Esther Perel, Paul Graham, Matt Mochary, Chris Camillo, Chris Voss, Stanley Druckenmiller, Jerry Colonna, Alex Hormozi, and Aakash Gupta.
Notable
The pipeline is grounded in an "11 elements of coaching" framework reverse-engineered from real coaching transcripts and ranked by leverage. Anti-patterns override LLM sycophancy ("never avoid hard truths" is the single highest-leverage line); example interactions teach behavior more reliably than description; pattern detection catches avoidance and false confidence in real time; and directness is earned through demonstrated understanding rather than asserted.
Stack
Status
Open source at github.com/blaizew/personas.