film-shot-prompt
Pass
Audited by Gen Agent Trust Hub on Sep 25, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTIONCOMMAND_EXECUTION
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill is designed to process external context from files such as
script.md,continuity.md, andasset-map.json. This ingestion of untrusted external data creates a theoretical surface for indirect prompt injection. However, the skill significantly reduces this risk by enforcing a strict, multi-layered template ('Prompt anatomy') and 25 'Hard rules' that scope the agent's output specifically to the cinematography domain. - Ingestion points: Reads
direction.md,script.md,continuity.md,asset-map.json, and user-provided image URLs. - Boundary markers: The skill requires a mandatory 'Global style header' and a fixed 'Prompt anatomy' order (Style, Context, References, Camera feel, Shots, etc.).
- Capability inventory: Generates text prompts and local
cohubCLI commands. - Sanitization: Enforces 'Positive language only' and specific technical parameters (lens mm, meter measurements, timing beats) which constrain the model's ability to follow instructions potentially hidden in the input text.
- [COMMAND_EXECUTION]: The skill generates
cohub generateCLI commands with parameters like duration, resolution, and image URLs. These commands are generated as output for the user to review or for project-specific automation and do not involve unauthorized system access or privilege escalation.
Audit Metadata