The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill establishes a workflow where an AI agent processes feedback (notes and pins) provided by editors through a website interface to generate pull requests. This creates an attack surface where an editor could potentially provide malicious instructions to influence the agent's code modifications.
Ingestion points: Notes and pins entered by editors on the live website or preview deployments.
Boundary markers: The documentation notes that the agent's report is rendered with raw HTML escaped, but does not explicitly define delimiters for the input notes to prevent instruction leakage.
Capability inventory: The AI agent (Claude Code) operates within a GitHub Actions environment with permissions to read/write repository contents, issues, and pull requests.
Sanitization: The skill focuses on providing precise context (file names, CSS paths) to the agent but does not detail specific sanitization of the natural language notes provided by users.
[EXTERNAL_DOWNLOADS]: The skill initiates the installation of the @nuthinking/site-edit package from the NPM registry. This is a vendor-owned package associated with the skill's author and is used to provide the widget and server handlers.
[COMMAND_EXECUTION]: The skill configures GitHub Actions workflows (claude.yml and preview-email.yml) that execute shell commands, install dependencies, and run scripts to manage the automated editing loop. These include operations like npm install, gh pr create, and running custom JavaScript scripts for generating preview emails.

site-edit