The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted user data (X posts, fragments, topics) to generate content. It mitigates risk through a mandatory evidence chain:
Ingestion points: User-provided X posts, topic fragments, or content fragments (Phase 0 in SKILL.md).
Boundary markers: The agent is instructed to extract a 'one-sentence core claim' and seek user confirmation before proceeding with content generation.
Capability inventory: The skill writes files to the local vault (_草稿/公众号短文/), executes the pbcopy command to copy text to the clipboard, and invokes image generation tools.
Sanitization: Content is filtered through a 'Compliance Red Line Table' and references specific safety documents (违禁词避雷表.md, 合规-大模型使用口径.md) to ensure output adheres to domestic regulations.
[COMMAND_EXECUTION]: The skill utilizes the pbcopy utility to facilitate the transfer of generated text to the system clipboard. This action is transparent, initiated as part of the primary content delivery workflow, and uses a standard system tool.
[DATA_EXPOSURE]: The skill reads and writes to specific project directories within the user's workspace for draft storage and compliance checking. These operations are limited to the intended workspace context.
[DYNAMIC_EXECUTION]: The skill follows secure practices by explicitly instructing the agent to read image generation credentials from environment variables and prohibiting them from being written to files.

zmm-post