The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill exhibits an indirect prompt injection attack surface as it is designed to ingest and process untrusted external data from the web (e.g., film research, practitioner interviews, and official film pages).
Ingestion points: External data enters the agent context via web browsing for film discovery, restoration notes, practitioner interviews, and reputable criticism, as specified in references/loop-playbook.md.
Boundary markers: The skill provides robust boundary markers by mandating the use of source-backed and inferred labels for all observations. SKILL.md explicitly states that "Evidence boundaries are mandatory."
Capability inventory: The skill possesses the capability to browse the web and write to local files, specifically updating the references/mechanism-bank.md and creating detailed tables in work/shot-breakdowns/.
Sanitization: The workflow requires that observations be categorized by confidence levels (verified, high-probability, low-confidence) and source status, ensuring that unverified or potentially malicious content from external sources is flagged and not promoted into the core mechanism bank without passing a quality gate.
[SAFE]: The skill adheres to security best practices by implementing a tiered confidence system for evidence and strictly limiting tool usage to localized project directories. No malicious patterns, obfuscation, or unauthorized network operations were detected.

film-breakdown-distiller