The Agent Skills Directory

[SAFE]: The skill is entirely instructional and does not include any executable scripts, tool invocations, or automated commands. The frontmatter explicitly sets disable-model-invocation: true, which restricts the agent from calling external tools.
[SAFE]: All external links point to legitimate and well-known research-related content, specifically the personal blog of Nicholas Carlini and the Latent Space publication. These resources are provided for context and analogy, not for code execution.
[SAFE]: No patterns of prompt injection, data exfiltration, or obfuscation were found. The instructions are clear and aligned with the stated purpose of research advising.
[SAFE]: The skill does not request access to sensitive file paths, environment variables, or credentials. It operates within the context of the user's provided research descriptions.

credo-taste