The Agent Skills Directory

[PROMPT_INJECTION]: Static analysis identified the phrase 'ignore previous instructions' within SKILL.md. Analysis confirms this is part of a defensive 'Untrusted-source discipline' section, instructing the agent to ignore any command-like text within ingested data, which is a security best practice for handling external content.
[COMMAND_EXECUTION]: The skill directs the agent to execute shell commands to resolve environment variables (printenv) and generate unique slugs for work directories using hashing utilities (sha256sum or shasum). These operations are used for legitimate configuration and state management purposes.
[EXTERNAL_DOWNLOADS]: The skill fetches documentation from user-supplied URLs to process them. It includes logic to handle various formats such as Markdown and PDF, and defines specific behavior for known publishers like Anthropic.
[PROMPT_INJECTION]: As the skill processes untrusted data from documentation websites, it has an inherent exposure to indirect prompt injection. The instructions mitigate this risk by explicitly telling subagents to treat all ingested content as data to be digested rather than commands to be followed.

docpage-digest