The Agent Skills Directory

[COMMAND_EXECUTION]: The skill uses local Node.js scripts to automate development tasks by spawning system processes. Specifically, runtime/serve.mjs can launch code editors to open files at specific line numbers, and runtime/shoot.mjs executes headless browser commands to capture UI screenshots. These executions are core to the skill's utility and include safeguards such as directory-restricted path resolution and argument escaping.
[INDIRECT_PROMPT_INJECTION]: The skill provides a mechanism for ingesting untrusted data in the form of reviewer feedback, which creates a surface for indirect prompt injection.
Ingestion points: Reviewer comments are gathered via runtime/html-mark.js and presented to the agent as instructions for code modification.
Boundary markers: The feedback format uses specific metadata tags to delineate reviewer notes from element selectors and snapshots.
Capability inventory: The agent possesses the ability to modify project source code and execute local scripts based on the ingested instructions.
Sanitization: There is no automated sanitization of the reviewer-provided text, meaning a malicious comment could attempt to manipulate the agent into performing unauthorized file or system operations.

html-prototype-build