The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: Fetches its own repository from GitHub (github.com/SkylarKitchen/skills.git) and installs standard development tools like playwright via the official npm registry. It also downloads a private copy of the chromium browser to support its plan-rendering functionality.
[COMMAND_EXECUTION]: Executes local shell commands and Node.js scripts to manage the lifecycle of plan generation, including environment setup, validation checks, and building the final HTML page. It supports specifying a custom Node.js execution command (stored in brief.json) to ensure compatibility in managed or virtualized environments.
[INDIRECT_PROMPT_INJECTION]: The skill ingests user-provided house descriptions which are processed by intake.md and then interpolated directly into the prompts for sub-agents responsible for drawing the plans. Maliciously crafted descriptions could attempt to influence the sub-agents' behavior.
Ingestion points: User descriptions of the house, lot, and budget are read into brief.json.
Boundary markers: Absent; the sub-agent dispatch templates interpolate user data without explicit delimiters or instructions to ignore embedded commands.
Capability inventory: The drawing sub-agents write JavaScript data files (oNN.js) which are subsequently evaluated by the controller process and included in the rendered HTML output.
Sanitization: No specific filtering or escaping of user input is performed before interpolation into sub-agent prompts.
[DYNAMIC_EXECUTION]: Evaluates JavaScript code generated by AI sub-agents using Node.js's vm.runInNewContext for validation and measurement purposes (wishes.cjs, check.cjs). While intended for processing plan data, the execution of LLM-generated code presents a risk surface if the generating agent is influenced by malicious input.

house-options