A Claude Code plugin with 12 slash commands covering the full agent product lifecycle - from "should this even be an agent?" through eval design to production safety. Grounded in primary sources from Anthropic, OpenAI, Karpathy, Hamel Husain, Linus Lee, and real production deployments.
Skill Claude CodeCodex
Generate a complete PM spec (PRD-style document) for an AI agent product. Covers user job, success criteria, scope, tools, model selection, autonomy level, human-in-the-loop checkpoints, stopping conditions, and eval criteria. Use after architecture-pattern to lock down the product definition.
Skill Claude CodeCodex
Review the UX of an AI agent product against modern patterns beyond chat. Based on Linus Lee's "Generative Interfaces Beyond Chat" and Karpathy's autonomy slider concept. Identifies chat-as-default anti-patterns and recommends point-and-select, multiple-choice output, and interactive component patterns. Use when…
Skill Claude CodeCodex
Recommend the right agent architecture pattern from Anthropic's 5 (prompt chaining, routing, parallelization, orchestrator-workers, evaluator-optimizer) plus OpenAI's manager and decentralized patterns. Use after qualify-agent confirms an agent is needed.
Skill Claude CodeCodex
Position a feature on the autonomy spectrum (suggestion → execution → full autonomy) and design the verification loop. Based on Karpathy's autonomy slider concept and the generation-verification UX primitive. Use when deciding how much human oversight an AI feature needs.
Skill Claude CodeCodex
Model the per-task and monthly cost of an agent product. Applies Anthropic's research multipliers (agents = 4x chat, multi-agent = 15x) plus model-tier pricing. Identifies cost optimization opportunities. Use BEFORE building to avoid surprise infrastructure bills.
Skill Claude CodeCodex
Design a complete evaluation plan for an AI agent product. Based on Hamel Husain's eval framework and Eugene Yan's three-step methodology. Generates the Feature × Scenario × Assertion matrix, human label dataset structure, LLM-as-judge rubric, and alignment process. Use BEFORE shipping any AI feature.
Skill Claude CodeCodex
Catalog the likely failure modes for a specific agent product and design mitigations for each. Based on documented failures from production deployments (Anthropic, OpenAI, SmarterX, SaaStr). Use before launch to anticipate what will go wrong.
Skill Claude CodeCodex
Design a layered guardrail stack for an AI agent product. Based on OpenAI's 7-layer guardrail framework plus mandatory human-in-the-loop triggers. Determines which guardrails to build, in what priority, with what tripwires. Use BEFORE shipping to avoid the SmarterX-class failures.
Skill Claude CodeCodex
Pre-launch checklist for shipping an AI agent product to production. Covers training loop design, ops ownership, infrastructure (rainbow deployments, checkpointing), observability, rollback plans, and on-call alerting. Based on Anthropic's multi-agent production lessons and Lenny/Lemkin's deployment experience.
Skill Claude CodeCodex
Determine whether a feature should be a deterministic system, a single LLM call, a workflow, an agent, or a multi-agent system. Based on Anthropic's "bias toward simplicity" and OpenAI's three agent triggers. Use BEFORE designing any AI feature to avoid over-engineering.
Skill Claude CodeCodex
Audit an agent system for the specific failure modes that caused real production disasters — permission scoping, environment separation, destructive action gates, and backup isolation. Based on the SmarterX database deletion case study and Anthropic's production lessons. Use BEFORE granting an agent production access.
Skill Claude CodeCodex
Review a proposed tool definition (name, description, schema) for Agent-Computer Interface (ACI) quality. Based on Anthropic's principle that tool design is UX design for agents. Use when defining tools for any AI agent system.