AI agents often write Claude skills that fail silently, according to AI LABS. The group reports that agent-written skills frequently omit setup prerequisites, clear boundaries on what the skill cannot do, and error-handling steps. Scripts inside those skills can also read secret files, expose an application to a local network, or point to a folder that exists only on one machine.1

Skill discovery creates a second bottleneck during real sessions, AI LABS says. Because AI agents see only a list of skill names and descriptions, long descriptions get truncated. In AI LABS's setup, descriptions were cut at 1,536 characters. With about 120 Claude Code skills installed, 74 appeared as names only, leaving agents without guidance on when to invoke them.1

SkillEvaluator is available for free on GitHub, and its security scan runs on SkillSpector, another NVIDIA AI tool, according to AI LABS. For deeper tiers, builders can use NVIDIA's free test key or supply their own key from Anthropic or OpenAI. To avoid a version mismatch that AI LABS hit when following the repository's own instructions, the group recommends pasting the GitHub link into an agent and asking it to install the tool.1

AI LABS tested two skills to show how the audit plays out in practice. Its AI slop detector scored a B (80.8), with its 2,150-character description cut off before the sentence explaining when to use it. A custom skill built over several rounds with Anthropic's skill-creator initially scored a B (82) with 49 problems. After revisions, it rose to an A (95.2), and its description shrank from 1,460 to 940 characters so the usage trigger appeared first.1

Builders should still review the tool's output carefully, AI LABS advises. Because many security alerts turned out to be false alarms, the group recommends walking through each finding with an agent before making changes. AI LABS also notes that running the same task with and without a skill is a closer test of real-world value than the layout score alone.1