The Agent Skills Directory

[PROMPT_INJECTION]: Susceptible to Indirect Prompt Injection. The skill ingests data from public web articles (via scrapers in a_stock_watcher/sources/) and passes this content directly to the Gemini AI model in a_stock_watcher/ai_parser.py. An attacker could publish content on the target website that includes malicious instructions designed to manipulate the AI's analysis or agent behavior.
Ingestion points: Scraped content from jiuyangongshe.com processed in a_stock_watcher/sources/study_hot.py, industry_chain.py, and action.py.
Boundary markers: None. The scraped article title and content are interpolated directly into the PARSE_PROMPT and IMAGE_PARSE_PROMPT without delimiters or instructions to ignore embedded commands.
Capability inventory: Network access for scraping, local SQLite database writes, and generation of natural language investment advice to the user.
Sanitization: No sanitization, filtering, or validation is performed on the scraped text before it is processed by the LLM.
[COMMAND_EXECUTION]: The skill frequently uses shell commands for environment setup, dependency management with uv, and execution of its internal Python modules and data query scripts.
[EXTERNAL_DOWNLOADS]: During the installation phase, the skill executes playwright install chromium, which downloads browser binaries from the official Playwright project servers to enable web scraping functionality.

jiucai-capture