The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes user-provided screenplays and external research data, which constitutes an indirect prompt injection attack surface. However, the skill includes explicit 'Trust and Authority' and 'Read-Only Trust Boundary' sections (in SKILL.md and references/screenplay-cold-read-protocol.md) that instruct the agent to treat all ingested content as untrusted data rather than instructions. * Ingestion points: User-supplied story material and external research results gathered via web search. * Boundary markers: Present. Instructions explicitly define a trust boundary and warn against following instructions embedded in analyzed text. * Capability inventory: The skill performs text analysis and generation; it does not request or use capabilities for arbitrary command execution or unauthorized network transmission. * Sanitization: Instructions mandate treating external content as evidence to be analyzed, specifically forbidding the execution of embedded requests or tool calls found in the data.
[EXTERNAL_DOWNLOADS]: The skill references several external URLs to GitHub repositories (e.g., VideoClaw, ViMax, genmedia-izumi-agent) and official film studio websites (e.g., NEON, Universal, Sony) in references/github-project-watchlist.md and references/screenplay-exemplar-benchmarks.md. These are used neutrally as sources for research, logic patterns, and creative benchmarks. The skill includes a 'Research Stop Rule' and requires user authorization for any updates based on this research.
[SAFE]: No signs of obfuscation, malicious persistence, credential theft, or unauthorized privilege escalation were detected. The skill's architecture is modular and focuses on creative logic within a controlled environment.

director-agent