The Agent Skills Directory

[COMMAND_EXECUTION]: The skill includes instructions in SKILL.md to run a local maintenance script (python3 scripts/selftest.py) to verify the integrity of asset data. The associated scripts (asset_check.py and selftest.py) were analyzed and found to contain only safe logic for parsing and validating JSONL records using standard Python libraries. No suspicious system calls or networking operations are present.
[INDIRECT_PROMPT_INJECTION]: The skill is designed to process untrusted drama scripts (剧本.md) which could theoretically contain instructions intended to influence the agent's behavior. However, the skill provides robust defense mechanisms through its structured workflow and validation requirements.
Ingestion points: Untrusted drama scripts and creator decisions enter the context through the extraction workflow.
Boundary markers: The instructions in references/occurrence-extraction.md and references/asset-review-checklist.md provide clear boundaries, requiring the agent to use 'minimal facts' and 'recognition anchors' rather than interpreting narrative subtext or ambiguous text.
Capability inventory: The agent is authorized to write to 视觉设定.md and perform structural analysis of the production data.
Sanitization: The skill includes a dedicated validation script (scripts/asset_check.py) to ensure all extracted data follows the required schema and maintains consistent identity bindings.

short-drama-assets