Five Rules for Evaluating Agent Skills Before They Bloat Your Context Window
Google, Tuesday, November 3rd, 2026
Google shares five rules for evaluating agent skills so SKILL.md files help rather than bloat an agent's context window.
Google describes lessons from measuring its Google Cloud Developer Plugin skills across Gemini, Antigravity and other agents in isolated sandboxes on GKE Autopilot.
Its first rule is that when a skill shows no improvement, the agent usually never loaded it, because routers only read the skill's name and description.
It warns of under-triggering and sibling hijacking, where a neighboring skill's broad description intercepts the prompt.
Google has open-sourced a tool called skill-reach that validates frontmatter, measures vocabulary overlap with rival skills and runs routing probes across Claude Code, Antigravity and OpenAI Codex.