The Agent Skills Directory

[COMMAND_EXECUTION]: The helper script scripts/find_refs.py makes extensive use of subprocess.run to invoke system utilities including curl, ffmpeg, and ffprobe. These tools are used to fetch design assets and generate contact sheets for the agent to review. While the script correctly passes arguments as lists to mitigate shell injection, the execution of complex media processing tools on untrusted external data carries a low-level risk of exploiting vulnerabilities in those parsers.
[EXTERNAL_DOWNLOADS]: The skill is designed to search and retrieve content from external sources including are.na, pinterest.com, sketchfab.com, and huggingface.co. It includes a generic get command that allows the agent to download files from any user-provided or discovered URL, placing them into the .context/refs/ directory.
[INDIRECT_PROMPT_INJECTION]: The skill's workflow involves ingesting metadata from third-party design platforms, creating a surface for indirect prompt injection.
Ingestion points: The scripts/find_refs.py script retrieves and prints block titles, descriptions, and source information from are.na and Pinterest. This text is then read by the agent to construct the "beats" and "look" sections of a prompt.
Capability inventory: The skill allows the agent to write files to the local filesystem, perform network requests via curl, and execute media processing commands.
Boundary markers: While the skill uses structured templates for its output, the instructions do not explicitly direct the agent to ignore instructions that might be maliciously embedded in the titles or descriptions of the retrieved references.
Sanitization: The script uses standard URL encoding for API queries and avoids shell interpolation for commands. However, the semantic content of retrieved metadata is not filtered for potential injection patterns before being processed by the agent.

blank-page