north-starr-genai

North Starr GenAI

Your North Starr for AI Development | v0.15.0

An agentic AI development agency framework — North Starr plans, designs, validates, and orchestrates while Claude Code writes code in YOUR codebase. Works with any project: RAG pipelines, agent harnesses, multi-agent systems, prompt chains, or AI platform components.

Agent Interaction Map

                    ┌──────────┐
                    │ /assess  │ ← classifies project type
                    └────┬─────┘
                         │
                    ┌────┴─────┐
                    │/discover │ ← elicits requirements (if needed)
                    └────┬─────┘
                         │
                    ┌────┴──────┐
                    │/decompose │ ← PRD → stories
                    └────┬──────┘
                         │
                  ┌──────┴──────┐
                  │ ORCHESTRATOR │
                  └──────┬──────┘
                         │
              ┌──────────┴──────────┐
              │    chief-ai-po      │ TRIAGE
              └──────────┬──────────┘
                         │
              ┌──────────┴──────────┐
              │    ai-architect     │ DESIGN
              └────┬──────────┬─────┘
                   │          │
         ┌─────────┘          └──────────┐
         ▼                               ▼
   ┌──────────┐                 ┌───────────────┐
   │ ai-invert│                 │cost-estimator │
   └────┬─────┘                 └───────┬───────┘
        └───────────────┬───────────────┘
                        ▼
                 ┌──────────────────┐
                 │ genai-layoutplan │ PLAN (tags tasks with specialists)
                 └────────┬─────────┘
                        │
       ┌────────────────┼────────────────┬──────────────┐
       ▼                ▼                ▼              ▼
┌────────────┐  ┌────────────┐  ┌───────────────┐ ┌──────────┐
│  prompt-   │  │   rag-     │  │ integration-  │ │ agentic- │
│  engineer  │  │  advisor   │  │   planner     │ │ designer │
└─────┬──────┘  └─────┬──────┘  └───────┬───────┘ └────┬─────┘
      └───────────────┼─────────────────┼───────────────┘
                      ▼ BUILD
        ┌─────────────┼──────────────┐
        ▼             ▼              ▼
┌────────────┐ ┌────────────┐ ┌───────────┐
│   eval-    │ │ guardrails-│ │  ai-ops   │
│  designer  │ │  designer  │ │ (monitor) │
└─────┬──────┘ └──────┬─────┘ └─────┬─────┘
      │          ┌────┴─────┐       │
      │          │  prompt- │       │
      │          │adversary │       │
      │          └──────────┘       │
      └───────────────┼────────────┘
                      ▼ HARDEN
               ┌────────────┐
               │demo-builder│ DELIVER
               └────────────┘

Feedback loops:
  eval fails ──→ prompt-engineer (fix prompt)
  guardrails fail ──→ ai-architect (fix design)
  cost overrun ──→ ai-architect (cheaper model)
  same gate fails twice ──→ HUMAN escalation

What It Does

North Starr GenAI is the brain of an AI development agency. It doesn't generate code — it generates the specs, designs, evaluations, and guardrails that make AI code production-grade.

CLIENT (you) → gives requirement
NORTH STARR (brain) → plans, designs, validates, orchestrates, quality-gates
CLAUDE CODE (hands) → reads North Starr's specs + writes code in YOUR codebase
/genai-bootstrap → makes Claude Code aware of your specific codebase patterns

The Pipeline

Requirement → /assess (classify project type)
           → /discover (elicit requirements if needed)
           → /decompose (PRD → stories with AI safety criteria)
           → /orchestrate (start the pipeline)
             → TRIAGE: chief-ai-po refines story
             → DESIGN: ai-architect → ADR + cost envelope
             → PLAN: genai-layoutplan → tasks with specialist tags
             → BUILD: specialists produce specs → Claude Code implements
             → HARDEN: eval + guardrails + ops validate (ALL must pass)
             → DELIVER: demo-builder packages for client

The AI Complexity Gate

Before ANY code change, the gate catches AI-specific risks:

#	Question	Why
Q0	Is current behavior covered by evals?	Eval-first discipline
Q1	Does this touch a production prompt or model config?	Prompt changes are high-risk
Q2	Does this change what data the model sees?	Data changes alter model behavior
Q3	Does this affect a client-facing output?	Client-visible changes need baselines
Q4	Could this change cost at scale?	Cost is a first-class concern

Based on answers, it routes through: ASSESS → BUILD (specialists auto-spawn) → HARDEN (validators auto-run) → COMPLETE → LEARN.

Routing Hooks (Claude Code only)

The plugin ships two hooks that fire automatically: