The Agent Skills Directory

[COMMAND_EXECUTION]: The script scripts/premiere_xml.py uses subprocess.run to execute ffprobe commands. The inputs for these commands, specifically file paths, are sourced directly from an external plan.json file. While it uses the list form of subprocess.run, there is no validation to ensure the paths do not contain malicious argument flags.
[DYNAMIC_EXECUTION]: The skill generates JavaScript code at runtime through scripts/figma_cards.py. This script interpolates data from a JSON configuration into a JavaScript template, writes it to a file, and the instructions then direct the agent to execute this generated code via the Figma MCP tool.
[INDIRECT_PROMPT_INJECTION]: The skill processes external plan.json files that contain potentially untrusted strings, such as marker text and file names. Ingestion points: The plan.json file processed by scripts/premiere_xml.py and referenced throughout SKILL.md serves as the primary entry point for untrusted data. Boundary markers: There are no explicit delimiters or instructions to the agent to ignore or treat the content of the JSON as untrusted data. Capability inventory: The skill has the capability to execute shell commands (subprocess.run), write files (XML and JS), and interact with professional software (Premiere Pro, Figma) via MCP tools. Sanitization: The provided scripts do not sanitize the text or path values from the JSON before they are used in command arguments or injected into the generated XML/JS files.
[REMOTE_CODE_EXECUTION]: The instructions in SKILL.md direct the agent to execute numerous 'project-side' Python scripts (e.g., aperta_final.py, aprende.py, fala.py, agente.py, rampa.py, onde.py, trecho.py, fonte_iphone.py, entrega.py, legenda_whisper.py, sensor_export.py) which are not included in the skill distribution. This encourages the execution of unverifiable local code.

visant-premiere-editor