The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted user-supplied prompts and multi-modal references, creating an ingestion surface for indirect instruction attacks. However, it incorporates significant mitigation and sanitization logic.
Ingestion points: SKILL.md workflow (Steps 1, 4, 11) and the Reference Role checklist in references/preflight-checklist.md process external user prompts and assets.
Boundary markers: The skill explicitly mandates that final model-ready prompts be isolated in a standalone 完整可复制提示词 (Complete Copyable Prompt) code block, separating generation-ready instructions from the agent's internal preflight reasoning.
Capability inventory: The skill is restricted to natural language analysis and local configuration reading; it possesses no network access or file-system write capabilities.
Sanitization: Implements a 'Positive-Only Context Rewrite Gate' and a 'Keyword Contamination Scan' to sanitize final prompts and remove adversarial or ambiguous phrasing.
[DATA_EXPOSURE]: The skill attempts to read workflow metadata and orchestration receipts from local relative paths (e.g., ../film-skill-orchestrator/references/mandatory-film-stack.md). This facilitates coordination within its intended multi-skill environment without exposing sensitive external data.
[SAFE]: The skill includes positive safety guardrails, such as a specific check prohibiting the use of adult body-proportion descriptors for characters that are school-aged or have ambiguous ages.

ai-video-prompt-preflight