The Agent Skills Directory

[PROMPT_INJECTION]: No evidence of prompt injection or safety bypass patterns was found. The skill actually promotes a structured prompt engineering methodology ('Universal Video Prompt Spec') that separates user intent from model dialect, which is a defensive technique against drift and injection.
[CREDENTIALS_SAFE]: The skill provides explicit security guidance regarding API keys. In the 'references/execution.md' and its Chinese counterpart, it instructs the agent to check environment variables (e.g., ATLASCLOUD_API_KEY) and strictly forbids asking users to paste keys into the conversation. It correctly distinguishes between environment-scope mismatches and missing keys.
[EXTERNAL_DOWNLOADS]: No suspicious external downloads or remote script executions were identified. The skill mentions fetching and polling from Atlas Cloud, which is the author's own infrastructure, and provides best practices for verifying downloaded media files to prevent processing corrupted data.
[COMMAND_EXECUTION]: The execution guidelines describe a read-only status-query state machine for monitoring generation tasks. It emphasizes using recorded IDs for polling rather than resubmitting tasks, which prevents accidental cost escalation and avoids dangerous shell command patterns.
[DATA_EXFILTRATION]: No patterns for sensitive data harvesting or unauthorized exfiltration were found. The skill focuses on legitimate task management and media generation workflows.
[NO_CODE]: The skill consists entirely of markdown documentation and instructional references; it does not contain or attempt to execute any scripts, binaries, or obfuscated payloads.

universal-video-prompt-skill