The Agent Skills Directory

[SAFE]: The skill serves as meta-documentation and a guide for skill engineering. It contains no executable scripts or commands that would pose a risk to the host system.
[EXTERNAL_DOWNLOADS]: The skill includes a reference inventory pointing to official documentation and technical guides from trusted organizations and well-known services, such as Anthropic, HumanLayer, and the Agent Skills Open Standard. These links are documented neutrally as resources for the user.
[COMMAND_EXECUTION]: While the documentation describes how to use shell commands (e.g., git status, chmod +x) and scripts within a skill, these are presented as educational examples for authors and are not executed by the skill itself in a malicious manner.
[PROMPT_INJECTION]: The instructions are designed to improve the clarity and reliability of agent behavior. There are no patterns indicative of bypassing safety guidelines or overriding system instructions.
[INDIRECT_PROMPT_INJECTION]: The skill defines an attack surface by instructions that analyze external skill files.
Ingestion points: Processes user-provided SKILL.md files during evaluation and debugging tasks.
Boundary markers: Explicitly recommends the use of XML tags and clear structural delimiters to prevent instruction confusion.
Capability inventory: Documents the use of file system access and shell execution tools as standard features for skill authors.
Sanitization: Provides rubrics for scoring and auditing skill quality, emphasizing structural integrity and clear instruction-following.

skill-engineering