The Agent Skills Directory

[SAFE]: The skill package consists exclusively of Markdown-based instructions and reference documentation. No executable scripts (.sh, .py, .js), binaries, or active code components are included or referenced for execution.
[OBFUSCATION]: The file .clawnet/unsigned-skill.json contains a hex-encoded string in the opReturnHex field. Decoding reveals a serialized version of the skill's own metadata and Markdown body. This is a platform-specific data integrity and registration mechanism (standard Bitcoin OP_RETURN format) rather than a malicious attempt to hide executable commands.
[INDIRECT_PROMPT_INJECTION]: The skill is designed to process untrusted human-facing prose as its primary input. While this represents an injection surface, the skill has no defined capabilities (no allowed-tools, no network access, and no file system write permissions) that could be exploited. Furthermore, the instructions include explicit 'closed-world constraints' and factual boundary checks that serve as safety delimiters for processing external data.
[EXTERNAL_DOWNLOADS]: Metadata in the benchmark files refers to a TypeScript script located on the author's GitHub repository. This resource belongs to the skill's vendor infrastructure and is used only for performance evaluation, not for runtime execution by the agent.

humanize