The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill analyzes user-supplied CAD files. This ingestion of external data represents a surface for indirect prompt injection. However, the risk is effectively managed because the analysis is conducted by a deterministic local script (scripts/mold_tool.py) that reports geometry measurements. The instructions explicitly direct the agent to rely on verifiable evidence rather than inventing data. Evidence chain: (1) Ingestion points: scripts/mold_tool.py reads mesh files via trimesh. (2) Boundary markers: SKILL.md mandates specific report formats and provenance for measurements. (3) Capability inventory: The skill is restricted to geometry math and inspection without network or writing permissions. (4) Sanitization: Input processing is handled by the established trimesh library using defined mesh formats.\n- [EXTERNAL_DOWNLOADS]: The skill references official manufacturing design guides from Protolabs and Hubs. These are well-known and trusted sources in the industry, and the references are provided for data lookup rather than for executing remote code.

dfm