north-starr-genai

Run AI-specific inversion analysis on a requirement before implementation. Dispatches the `ai-invert-analyst` agent on a separate thread. Use before complex or high-stakes AI tasks that touch prompts, models, RAG, or AI-powered outputs.

Autonomously improve any skill or agent prompt via measure-change-test hill-climbing. Dispatches the `auto-improver` agent on a separate thread. Use when a skill gives inconsistent results, when asked to "improve/optimize/autoresearch" a skill, or when output quality needs iterative tightening.

Capture a reproducible AI performance baseline before changes. Dispatches the `baseline-capturer` agent on a separate thread. Use before any change to a client-facing AI output. Triggered automatically by the Q3 gate in CLAUDE.md.

Generate AI tool configuration for an existing project. Explores the codebase and produces context files, path-scoped pattern rules, landmine rules, and agents for Claude Code and VS Code Copilot. Run once per project.

Inject managed sections into existing CLAUDE.md and AGENTS.md after a North Starr GenAI plugin update without re-bootstrapping. Preserves all project-specific content.

Generate executable pytest test files for AI outputs. Produces assertion-based tests for deterministic AI components (classification, extraction, routing, structured output) that run in CI/CD. Complements /eval-suite which produces statistical evaluation datasets for non-deterministic outputs.

Pre-deployment verification checklist for AI automations. Verifies evals pass, baselines show no regressions, guardrails are tested, costs are reviewed, monitoring is configured, and rollback plan is documented.

Deep structured inversion analysis on a requirement before implementation. Identifies risks, edge cases, failure modes, and convention violations. Use before complex or high-stakes tasks.

Generate input/output guardrail specifications for an AI automation. Covers PII detection, prompt injection filtering, content safety, confidence thresholds, human escalation, fallback behavior, and audit logging.

Generate client-facing documentation for completed AI automations. Includes system overview, architecture diagram, monitoring guide, prompt modification guide, escalation procedures, and SLA summary.

Generate runbooks for common AI automation failures. Covers model timeouts, hallucination detection, cost spikes, accuracy drops, data pipeline failures, and prompt injection attempts.

Track prompt iterations with diffs, scores, and rollback capability. Version control for prompts — the "source code" of AI automations.

Decompose a PRD into prioritized, dependency-mapped epics and user stories. Accepts pasted text, file path, or PDF.

Start the AI story orchestration pipeline. Feeds stories from a story map into the orchestrator agent, which routes them through TRIAGE → DESIGN → PLAN → BUILD → HARDEN → DELIVER with feedback loops, conflict detection, and human escalation.

Generate evaluation datasets from requirements. Creates golden examples, adversarial inputs, boundary cases, and regression anchors with scoring rubrics. Use when building test suites for AI outputs.

Classify project type, recommend approach, identify needed agents, estimate complexity, and flag risks. Runs BEFORE /decompose to help North Starr adapt its pipeline to what is being built.

Elicit requirements from a client who has a problem but no PRD. Asks structured questions, identifies actual needs, and generates a PRD draft ready for /decompose.

Estimate token costs for a proposed AI automation or change. Projects per-request and monthly costs at 1x, 10x, 100x scale with model comparison. Use before building or when the complexity gate detects cost impact (Q4=Yes).

After completing a task, capture learnings by updating AI tool configuration. Creates or updates rules, agents, and context files based on what was discovered — in the native format for each tool (Claude Code, VS Code Copilot, AGENTS.md). Use after any task where something surprised you, went wrong, or produced a reusable insight.

Run a prompt against a test suite of inputs, score outputs, and compare against baseline. Use for single-run prompt evaluation after changes.