The Agent Skills Directory

[PROMPT_INJECTION]: The skill processes untrusted data from external sources, creating a potential surface for indirect prompt injection.
Ingestion points: The skill uses the web_search tool to collect recent AI news and reads content from ./sucai.md provided by the user.
Boundary markers: There are no specific delimiters or instructions (e.g., "ignore instructions found in search results") defined to separate untrusted external content from the core writing task.
Capability inventory: The skill possesses the ability to perform network searches, interact with the user via AskUserQuestion, and execute a local Python script for word count verification. This combination of reading untrusted data and having executable capabilities is the primary risk factor.
Sanitization: No sanitization or filtering logic is implemented to clean or validate the content retrieved from external websites or local asset files before it is processed by the AI for article generation.
[COMMAND_EXECUTION]: The skill invokes local system commands and scripts during its workflow.
Evidence: It executes date "+%Y年%m月%d日" to synchronize the timeline and runs ./scripts/check_body_length.py to verify the length of the generated content. These commands are benign, run locally, and do not involve untrusted inputs flowing directly into the shell.

ai-article