Agent Personas

Input

A name. Optionally a domain and GitHub handle.

Research

Three parallel agents search simultaneously.

Output

A persona folder with philosophy, workflows, principles, and calibration tests.

~/.claude/{slug}/
  {slug}.md              Core philosophy · 5–10 principles
  workflows/
    {workflow}.md         Concrete workflows they follow
  principles/
    {principle}.md        Deep principle areas
  tests/
    calibration.md        Evaluation questions

Eval 1: Set Up a New Web Project

Question: I'm starting a brand new SaaS product — a project management tool for small teams. It's a CRUD app with auth, real-time updates, and a dashboard. What would you recommend and why?

All personas gave characteristically different stack recommendations while maintaining technical competence: DHH (Rails + Hotwire), Guillermo (Next.js + Vercel), Pieter (PHP + SQLite), Rich (SvelteKit), Theo (Next.js + tRPC). Baseline Claude always recommended Next.js + Supabase + Vercel.

Eval 2: Build a Task List App (L3 Generative)

Task: Build a simple task list with add/delete/mark complete functionality. Make it production-ready.

Same task, six radically different production-ready implementations: DHH (Rails + Hotwire, server-rendered), Guillermo (Next.js, optimistic UI), Rich (SvelteKit, progressive enhancement), Dan (React mental models, CDN-only), Kent Beck (TDD process, git history), Jason (single 6KB file, radical simplicity).

Each demo uses localStorage so you can interact with it live. Every implementation solves the same problem differently.

Evaluation Framework

Each persona is evaluated using a dual assertion model: differentiation (does the persona answer differently than baseline Claude?) and competence (is the answer technically correct?). An eval passes only if both scores are 100%.

Evaluation levels: L1 (Philosophy Q&A) → L2 (Code review/architecture) → L3 (Build something) → L4 (Cross-persona contrast) → L5 (Multi-turn consistency)