The Agent Skills Directory

[DYNAMIC_EXECUTION]: The execution entry script scripts/run.py uses os.execv to perform process replacement. It dynamically determines the path to a virtual environment interpreter within a hidden directory (.runners-venv) and re-executes itself with that interpreter. This constitutes execution of code from a computed filesystem path.
[COMMAND_EXECUTION]: The skill orchestrates several lower-level tools and CLI runners, including common.runners.cli.carousel and common.runners.cli.reel, through Python module execution and shell commands. These calls use relative paths within the skill's directory structure.
[INDIRECT_PROMPT_INJECTION]: The skill is designed to ingest and process potentially untrusted external data from user-supplied research briefs and content files, which are then used to build prompts for image-generation sub-agents.
Ingestion points: Input data is accepted via the --research and --content-file command-line arguments, which point to local markdown or text files.
Boundary markers: The system relies on a central SYSTEM_PROMPT and structural constraints (such as JSON output requirements) to maintain control over the sub-agent's behavior.
Capability inventory: The skill environment allows for filesystem modifications, execution of shell commands, and data uploads to S3, which could be leveraged if the prompt generation logic is successfully subverted by malicious input data.
Sanitization: Input content is filtered and mapped into specific structural 'roles' for each slide, providing a level of abstraction that limits the direct flow of raw input into the generation prompt.

carousel-builder