The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill fetches content from the web to research topics for video creation, creating a potential surface for indirect prompt injection.
Ingestion points: SKILL.md Step 2 (WebSearch and WebFetch results used to populate the storyboard).
Boundary markers: The skill explicitly instructs the agent in Step 2 and the Constraints section to treat all fetched text as untrusted data and discard any embedded instructions (e.g., "ignore previous instructions").
Capability inventory: The skill utilizes npx remotion render for subprocess execution and writes to props.json for data storage.
Sanitization: The agent-generated storyboard is validated against a strict Zod schema in project/src/schema.ts within the Remotion project, ensuring only valid data structures are processed by the React-based renderer.
[COMMAND_EXECUTION]: The skill uses shell commands to process filenames and execute the video rendering toolchain.
Evidence: SKILL.md Step 4 demonstrates the use of a pipeline (echo, tr, sed, cut) to sanitize user-supplied briefs into safe filename slugs.
The rendering command npx --no-install remotion render is a standard use of the Remotion CLI for local rendering within the CI environment.
[PROMPT_INJECTION]: While static analysis flagged potential prompt injection patterns in SKILL.md, these are confirmed to be defensive instructions directing the agent to ignore malicious prompts found in external data rather than attempts to subvert system guidelines.

remotion