The Agent Skills Directory

[DATA_EXFILTRATION]: The skill reads the WRITING_MODEL_API_KEY from a local environment file and transmits it, along with article content, to a user-configured base_url. While this is documented as core functionality, it creates a risk if the user points the tool to an untrusted or malicious API endpoint.
[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted user-provided content from topic cards and business reference documents, which are directly interpolated into LLM prompts.
Ingestion points: topic-card.md and markdown files located in .aws-article/products/.
Boundary markers: The skill uses markdown headers like '选题卡片' and '参考资料库' to separate input data in the prompt.
Capability inventory: The system allows writing to draft.md and making network requests to remote API endpoints.
Sanitization: There is no evidence of filtering or validation of the input text to prevent malicious instructions from influencing the AI's behavior.
[COMMAND_EXECUTION]: The skill executes a local Python script scripts/write.py via the python3 interpreter to manage writing logic, prompt assembly, and API communications.

aws-wechat-article-writing