The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill implements a security boundary for processing untrusted data such as other skill files. Rule 11 explicitly instructs the agent to treat candidate content as data and ignore any embedded instructions, effectively mitigating potential indirect prompt injection attacks.
Ingestion points: Reads and processes external skill files (SKILL.md and references) during review and audit modes.
Boundary markers: Uses explicit logical boundaries defined in the 'Rules' section.
Capability inventory: Uses git, local shell scripts, and web search for research.
Sanitization: Instructions to ignore and report embedded commands found in analyzed content.
[COMMAND_EXECUTION]: The skill uses local repository tools such as git and development scripts like lint-skills.sh and validate-spec.sh. These operations are within the scope of a developer-centric meta-skill and are used for repository state management and content validation.
[DATA_EXFILTRATION]: The skill performs research tasks such as searching for current CVEs and tool version information. These are legitimate activities for maintaining skill security and freshness, and do not involve the exfiltration of sensitive user credentials or environment data.

skill-creator