The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted data from two main ingestion points: OCR text from user-uploaded images and content retrieved from the web via search tools. This creates a surface for indirect prompt injection where adversarial content could influence the agent's analysis.
Ingestion points: User screenshots (K-line and financial reports) and web search results (news, analyst reports, and industry trends).
Boundary markers: The instructions do not define explicit delimiters or include safety warnings to ignore instructions embedded in the external data.
Capability inventory: The skill possesses file-write capabilities (~/stock-reports/), subprocess execution for local Python scripts, and network access for search and document uploading.
Sanitization: There is no evidence of sanitization or filtering of external data before it is interpolated into the analysis prompts.
[COMMAND_EXECUTION]: The workflow involves executing local Python scripts via shell commands with arguments (like stock symbols) that originate from user input or OCR results. This represents a potential command injection vector if the agent's execution environment does not properly escape these parameters.
[EXTERNAL_DOWNLOADS]: The skill utilizes the akshare library to fetch financial data from external providers and uses playwright for PDF generation, which may load external resources such as fonts or stylesheets during the rendering process. These are legitimate operations consistent with the skill's purpose.

stock-advisor