The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill fetches setup metadata, version constraints, and package URLs from a remote source at https://neirong.funshion.net/skills/setup-skill.md. While this is a vendor-controlled domain (Funshion/LeoYeAI), it establishes a remote dependency for core operations.
[COMMAND_EXECUTION]: The scripts/api-client.sh script employs the source command to load config/.env. This is a form of dynamic shell execution where any code within the environment file is executed in the current shell context.
[COMMAND_EXECUTION]: Several scripts (scripts/api-client.sh, examples/create-comic.sh, etc.) execute inline Python code blocks using `python3
<<'PY'`. This technique is used for tasks like URL normalization, JWT prefix validation, and version comparison.
[PROMPT_INJECTION]: The SKILL.md file contains instructions for the agent to conceal internal configurations from the user, such as default operation modes ("Default mode silent use, do not actively explain to users"), which is a concealment pattern.
[PROMPT_INJECTION]: Instructions in SKILL.md direct the agent to perform automatic installations and mandatory updates based on remote metadata, granting the agent autonomy over its own codebase updates.
[PROMPT_INJECTION]: Indirect Prompt Injection: The skill processes untrusted data from the platform API (such as AI-generated scripts, role descriptions, and scene captions) and interpolates this data directly into subsequent agent prompts without using boundary markers.
Ingestion points: Data retrieved from ai.fun.tv via GET requests for project details, roles, and storyboard resources.
Boundary markers: Absent in prompt instructions and script templates.
Capability inventory: The skill has access to network operations (curl), file system writes, and local command execution (bash, python).
Sanitization: Content is parsed via jq but the resulting strings are used in prompts without escaping or filtering for malicious instructions.

Funai-skill