ai-engineering-lifecycle

Create representative golden, adversarial, safety, and failure evaluation cases, then measure quality, groundedness, safety, latency, reliability, and cost against a non-LLM baseline. This Skill walks through that evaluation process step by step.

Apply prompt-injection defenses together with least-privilege tools, data isolation, approval boundaries, audit logging, and fallback behavior. The lifecycle dedicates a full step to these security controls before deployment.

Select retrieval, agent or workflow patterns, model routing, memory, and hosting only through an explicit RFC decision after defining the user outcome and data contract. The Skill enforces this decision gate rather than defaulting to a pattern.

No. It targets LLM and agentic application engineering rather than model-training research, focusing on application-level concerns like evals, tools, observability, and deployment.

Instrument inputs, outputs, tool calls, retrieval, evals, errors, latency, cost, and version identifiers while avoiding leakage of sensitive data. This gives traceability for debugging and production-readiness review.