The Agent Skills Directory

[PROMPT_INJECTION]: The skill instructions are purely diagnostic and instructional, focusing on visual auditing. No attempts to override safety filters, bypass system prompts, or use 'DAN'-style jailbreak commands were found in the markdown body or test cases.
[DATA_EXPOSURE]: No sensitive file paths (e.g., .ssh, .env), hardcoded credentials, or API keys were detected. The skill does not request or perform network operations that could facilitate data exfiltration.
[OBFUSCATION]: The content is written in clear Chinese and English prose. No Base64-encoded payloads, zero-width characters, homoglyph attacks, or hidden URL patterns were identified.
[REMOTE_CODE_EXECUTION]: The skill contains no executable scripts, shell commands, or instructions to download and run external code. All test cases are localized to prompt-trigger logic.
[COMMAND_EXECUTION]: There are no instances of subprocess spawning, use of dangerous CLI tools, or dynamic context injection patterns (e.g., !command placeholders).
[INDIRECT_PROMPT_INJECTION]: While the skill is designed to process external visual sequences and shot descriptions, it lacks any high-privilege capabilities (like file-writing or network access) that could be abused if malicious instructions were embedded in the ingested data. The workflow is strictly analytical.
[METADATA_POISONING]: The YAML frontmatter and skill description accurately reflect the purpose of the tool without deceptive instructions or hidden malicious metadata.

cross-shot-consistency-audit