The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill ingests data from external zoning analysis reports and project context files to generate its 3D visualization. While it includes basic validation for coordinates and heights, the inclusion of untrusted data into a generated HTML file creates a surface for indirect prompt injection if the source data is maliciously crafted.
Ingestion points: zoning-analysis-*.md (JSON data block), PROJECT.md.
Boundary markers: No explicit delimiters or isolation warnings are used for the input data.
Capability inventory: Read, Write, Bash, Glob, Grep.
Sanitization: Instructions specify parsing JSON without executable evaluation and validating that geometric inputs are finite numbers.
[EXTERNAL_DOWNLOADS]: The generated HTML viewer fetches the Three.js library from a public CDN to render the 3D scene.
Source: https://cdn.jsdelivr.net/npm/three@0.170.0/.
Usage: Integrated via an HTML import map within the generated viewer.
[COMMAND_EXECUTION]: The skill uses shell commands (Bash, Glob, Grep) to search for and select zoning analysis reports in the current directory. It also executes a command to open the resulting HTML file in a web browser.
[DYNAMIC_EXECUTION]: The skill dynamically generates a JavaScript-heavy HTML file at runtime based on the contents of the analysis report.
Implementation: The agent assembles Three.js scene logic, utility functions, and geometry data into a standalone HTML artifact.
Execution: The resulting file is executed in the user's browser environment.

zoning-envelope