The Agent Skills Directory

[COMMAND_EXECUTION]: The skill instructions in SKILL.md describe executing local shell commands: node scripts/model-consult.js consult <model_id> --seat --prompt "…". The prompt parameter is intended to be filled with user-supplied creative briefs and existing screenplay material. This pattern poses a risk of command injection if the arguments are not correctly sanitized by the script or the shell.
[DYNAMIC_EXECUTION]: The skill uses Node.js scripts (scripts/model-consult.js, scripts/model-consult-mcp.mjs) to dynamically manage model routing and agent consultations at runtime. This adds complexity and execution paths that are not fully visible in the markdown instructions.
[INDIRECT_PROMPT_INJECTION]: The workflow facilitates the ingestion of external data such as "V1 script, source material, notes" which are then distributed as prompts to six specialized agents. 1. Ingestion points: Step 1: Brief the Room in SKILL.md accepts external scripts and notes. 2. Boundary markers: No specific delimiters or safety instructions are mentioned to prevent the LLM from following instructions embedded within the ingested scripts. 3. Capability inventory: The skill possesses the ability to execute shell commands (node) and communicate with external LLM providers. 4. Sanitization: There is no evidence of sanitization or filtering of the external screenplay content before it is interpolated into agent prompts.

writers-room