The Agent Skills Directory

[SAFE]: No malicious code or suspicious behavior detected. The skill follows security best practices for prompt engineering and data handling.
[INDIRECT_PROMPT_INJECTION]: The skill identifies an attack surface where untrusted data (e.g., DISCOURSE.md, BRAND.md, and competitor web search results) is ingested into the agent context. It mitigates this risk by explicitly instructing the agent to ignore embedded instructions within these sources and to use a fenced helper script for loading.
Ingestion points: External competitor content from web searches and project-level files (DISCOURSE.md, BRAND.md, VOICE.md) that are not part of the skill's own code.
Boundary markers: Explicit instructions to "ignore embedded instructions" and treat content within the "fenced helper" as untrusted.
Capability inventory: The skill uses web search and local file read operations; it does not contain subprocess execution or network write capabilities.
Sanitization: Instructs the agent to validate source URLs and surface warnings if the loading helper fails.
[DATA_EXPOSURE_AND_EXFILTRATION]: The skill references local files using paths relative to ${CLAUDE_SKILL_DIR}. It includes security constraints to refuse non-absolute overrides and resolve paths through a trusted mechanism, preventing path traversal or unauthorized access to sensitive system files.

blog-strategy