The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill has an inherent attack surface because it ingests untrusted third-party data from spreadsheets, CSVs, PDFs, documents, and presentations. This is proactively addressed through the 'Authority Boundary' section, which instructs the agent to ignore any instructions in the source documents that attempt to authorize tool use or API calls. Furthermore, a human-in-the-loop 'Approval Gate' is required after the planning phase, ensuring that no mutations occur without explicit user verification. The ingestion process is structured via a canonical ledger and specialized host readers to maintain data integrity.
[COMMAND_EXECUTION]: Although the skill is configured with the Bash tool, its instructions are limited to using host-provided file readers and delegating implementation logic to other approved internal skills. There are no patterns of arbitrary or dangerous command execution.
[DATA_EXPOSURE]: The skill performs file reads and network operations strictly within the context of media planning for the 'spotify' vendor. It interacts with the Spotify Ads API using an MCP server or a shared request wrapper and does not access sensitive local credentials (e.g., SSH keys, AWS configs) or exfiltrate data to unauthorized third-party domains.

media-plan-to-draft