The Agent Skills Directory

[REMOTE_CODE_EXECUTION]: The skill provides explicit instructions in references/generative-tools.md to download and build a third-party project from source. Specifically, it directs the agent/user to clone https://github.com/jamiepine/voicebox.git and run make setup && make dev. Downloading and executing build scripts from an unverified personal repository is a high-risk activity that can lead to arbitrary code execution on the host system.
[EXTERNAL_DOWNLOADS]: The skill references and encourages the download of several external tools. While some are from well-known services, the inclusion of a dependency from a personal GitHub repository (jamiepine/voicebox) for voice cloning introduces significant supply chain risk.
[COMMAND_EXECUTION]: The skill utilizes a variety of command-line tools for its operations. SKILL.md describes a workflow involving node to run internal ad platform scripts (e.g., tools/clis/google-ads.js), while references/generative-tools.md includes commands for curl, ffmpeg (for media processing), and npx (for rendering video ads with Remotion).
[PROMPT_INJECTION]: The skill is susceptible to indirect prompt injection through its data processing workflow. In 'Mode 2: Iterate from Performance Data', it ingests performance data from CSVs, pasted text, or API outputs. Because the agent has the capability to execute shell commands and interact with advertising platform APIs, malicious instructions embedded within the performance data could be used to manipulate the agent into performing unauthorized actions.
Ingestion points: Performance data processed in SKILL.md via CSV, text input, or tool-generated API output.
Boundary markers: There are no delimiters or specific instructions provided to ensure the agent ignores command-like text within the ingested performance data.
Capability inventory: Extensive access to advertising platform CLIs (google-ads, meta-ads, linkedin-ads, tiktok-ads), network tools (curl), and media processing utilities.
Sanitization: The instructions do not define any sanitization or validation steps for external performance data.

ad-creative