The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill ingests untrusted data from user-provided text and images (geometry problems) to derive structured specifications. These specifications are then used to drive logic in Python scripts and generate HTML output. While this creates a potential surface for indirect injection if a problem description contains adversarial instructions, the risk is mitigated by the specific mathematical domain and the use of structured JSON for data injection.
Ingestion points: Geometry problem descriptions provided as text or images (SKILL.md Step 1).
Boundary markers: Absent for the processed problem content.
Capability inventory: File system write access to the current working directory and execution of local Python scripts (SKILL.md Step 3).
Sanitization: The generate.py script uses json.dumps() to safely escape data injected into the HTML template, though it lacks path sanitization for the output file destination.
[EXTERNAL_DOWNLOADS]: The HTML template (template/lesson.html) loads standard web libraries, specifically Three.js from unpkg.com and MathJax from cdnjs.cloudflare.com. These are well-known, reputable services used for 3D rendering and mathematical typesetting.
[COMMAND_EXECUTION]: The skill requires the execution of Python scripts (geometry_kernel.py and generate.py) to perform symbolic math calculations using the sympy library. The instructions explicitly warn the agent not to install the sympy dependency without first obtaining user permission.

edu-solid-geometry