The Agent Skills Directory

[DATA_EXFILTRATION]: The script scripts/generate_script.py is vulnerable to path traversal. The project command-line argument is used to construct file paths (PROJECT_ROOT / "projects" / args.project) without any validation or sanitization. An attacker could potentially use .. sequences to attempt to access directories outside of the intended projects/ folder, although the script's requirement for a specific sub-folder structure (drafts/episode_N/step1_segments.md) significantly limits the practical impact.
[INDIRECT_PROMPT_INJECTION]: The skill represents a surface for indirect prompt injection as it ingests data from external files and incorporates them into prompts for the Gemini API.
Ingestion points: The skill reads content from step1_segments.md and project.json located within the project directory.
Boundary markers: There are no visible delimiters or 'ignore' instructions in the skill logic to prevent the LLM from following instructions embedded within these files.
Capability inventory: The skill has the capability to read files and invoke an external LLM (Gemini API).
Sanitization: There is no evidence of sanitization or filtering of the content read from the markdown and JSON files before they are processed by the generator.

generate-script