The Agent Skills Directory

[PROMPT_INJECTION]: The skill is susceptible to indirect prompt injection because it ingests untrusted data in the form of performance metrics (CSV, paste, or API output) to iterate on ad copy.
Ingestion points: Performance data provided by the user in SKILL.md (Mode 2: Iterate from Performance Data).
Boundary markers: Absent; there are no specific instructions to the agent to treat this data as untrusted or to use delimiters.
Capability inventory: The agent can write files, execute local CLI tools, and make network requests to various AI service APIs.
Sanitization: Not present; the data is directly analyzed to identify winning patterns and generate new content.
[REMOTE_CODE_EXECUTION]: The skill encourages the generation and execution of dynamic code. In references/generative-tools.md, it suggests using AI to design React templates for the Remotion framework, which are subsequently rendered into video files.
[COMMAND_EXECUTION]: The skill provides instructions and examples for executing shell commands and local scripts, such as node tools/clis/google-ads.js for data retrieval and npx remotion render for video production.
[EXTERNAL_DOWNLOADS]: The skill references external resources and APIs. It includes instructions to clone the Voicebox repository from GitHub (jamiepine/voicebox) and provides curl examples for interacting with official APIs from Google (Gemini) and ElevenLabs. These are well-known services used for their primary intended purpose.

ad-creative