The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill references the Dagre layout library from the well-known jsDelivr CDN (cdn.jsdelivr.net) within its HTML preview template. It also performs a standard connectivity check using curl to verify availability of the library.
[COMMAND_EXECUTION]: The skill uses common shell utilities for workspace management, including mkdir for directory setup, cp for template copying, and python3 -m http.server to serve preview pages locally. It also employs rg (ripgrep) and git for project validation.
[PROMPT_INJECTION]: The skill has an indirect prompt injection surface as it processes external research files and user feedback. Evidence of this includes: (1) Ingestion points: The skill reads project research/ files and user feedback nodes. (2) Boundary markers: Iteration IDs and node labels are used to structure data, though explicit prompt delimiters for external text are absent. (3) Capability inventory: The agent has shell execution and file-write capabilities across the workspace. (4) Sanitization: The skill addresses this by instructing the agent to "Record a redacted feedback summary" and use "safe summaries" in logs to prevent raw feedback from overriding agent instructions.

art-director