The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill ingests master script JSON files that are interpolated directly into the user prompt, creating an attack surface for indirect prompt injection. If an input script contains adversarial instructions (e.g., in a section body), the model might follow those instructions instead of the skill's defined rules.
Ingestion points: In scripts/lib/run.mjs, the tool reads external files via the --script and --brief (with @ prefix) arguments.
Boundary markers: The SOCIAL_CLIP_USER_PROMPT in scripts/lib/prompts.mjs does not use clear delimiters (such as XML tags, triple backticks, or specific separators) to isolate the untrusted script text from the system instructions. It also lacks a directive for the model to ignore any instructions found within the script content.
Capability inventory: The skill script scripts/lib/run.mjs has the capability to write files to the local filesystem (writeFileSync) and perform network requests to the OpenRouter API.
Sanitization: There is no evidence of sanitization, filtering, or escaping of the script text before it is joined and passed to the model in scripts/lib/prompts.mjs.

studio-clips