Google Cloud has been measuring and optimizing its Google Cloud Developer Plugin, running dozens of scenarios with Gemini, Antigravity, and other agents, and using Claude Sonnet 5 on the Gemini platform inside isolated Harbor pods on GKE Autopilot, while contributing Kubernetes and Gemini platform capabilities upstream to NVIDIA's open-source evaluator. The company outlined a five-rule engineering playbook for verifying a skill is worth its tokens. It warned that when an evaluation run shows zero improvement between plugin and baseline arms (Δ = 0), the agent likely never loaded the skill, since coding agents inspect only the short name and description in YAML frontmatter at turn zero; past a dozen skills, under-triggering and sibling hijacking can occur. Google open-sourced skill-reach to catch frontmatter collisions before spending on container sandboxes, validating manifests, computing vocabulary overlap against nearest rival skills, and running routing probes across Claude Code, Google Antigravity, and OpenAI Codex, paired with NVIDIA's SkillSpector. The playbook also warns against running skill benchmarks on developer laptops, where cached Google Cloud credentials, exported project variables and warm bash history, global environment rules, and host resource contention can contaminate the baseline arm and mask failures.