# Agent Starter Kit

A model-agnostic, dependency-free scaffold of Markdown personas and rules for coding agents, where a Maestro orchestrator breaks a plain-language request into sub-agent work handled by an Architect, Coder, Reviewer and Contextualizer, each able to run on a different model.

Agent Starter Kit is a Natural Language AI Harness: a foundation of Markdown personas, playbooks and rules that a coding agent such as Claude Code loads to work on a software project. A Maestro persona receives a plain-language request, breaks it down, and dispatches the work to specialised personas: an Architect who designs, a Coder who implements, a Reviewer who checks the result, and a Contextualizer who documents the project's structure.

Each persona defines its own identity, playbook, handoff format and rules for what it must not do, plus a self-review rubric it applies before delivering its work. A preferredModel setting routes each persona to a chosen model provider, so orchestration, planning and review can run on different models, or all on one if preferred. Corrections and lessons learned are captured to long-term memory and carried into future sessions, and interrupted work is tracked so a new session can resume it.

It is pure Markdown with no external dependencies, and is meant to be cloned into a project's own .agents/ directory and adapted per project.

Website: https://github.com/ntorga/agent-starter-kit