The Agent Skills Directory

[REMOTE_CODE_EXECUTION]: The installation section recommends a highly insecure pattern for acquiring the skill: curl -sSL https://raw.githubusercontent.com/cognyai/claude-code-marketing-skills/main/install.sh | bash. This allows for arbitrary command execution on the host system from a remote source that is not verified or listed as a trusted vendor.
[METADATA_POISONING]: There is a significant discrepancy between the system-provided author context ('reason-machines') and the entities referenced in the skill content ('ara.so' and 'cognyai'). This identity mismatch is often associated with supply chain attacks or unauthorized redistribution of code.
[INDIRECT_PROMPT_INJECTION]: The skill provides numerous functions (e.g., /seo-audit, /landing-page-review, /competitor-analysis) that ingest content from arbitrary third-party URLs using WebFetch and WebSearch tools. This creates an attack surface where malicious instructions hidden on those pages (e.g., in HTML comments or metadata) could influence the agent's behavior.
Ingestion points: /seo-audit, /landing-page-review, and /competitor-analysis commands in SKILL.md.
Boundary markers: None identified; there are no instructions to the agent to treat fetched web content as untrusted data.
Capability inventory: The skill documentation mentions interaction with various APIs (Google Ads, Meta Ads, LinkedIn Ads, HubSpot) and local file system access for installation.
Sanitization: None identified for the ingested web content.
[COMMAND_EXECUTION]: The documentation provides instructions for the agent to modify the local environment, including deleting cache directories (rm -rf ~/.cache/claude/mcp/cogny) and modifying MCP configuration files (.mcp.json).

claude-code-marketing-skills