The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill is designed to ingest and analyze HTML content from external, user-specified URLs, creating a surface for indirect prompt injection.
Ingestion points: The scripts/lib/pages.mjs script fetches HTML content from remote origins using the fetch API, and scripts/observe-runtime.mjs captures the live DOM from a headless browser.
Boundary markers: The instructions do not define specific delimiters or instructions to the model to ignore potential malicious content embedded within the analyzed website's markup (e.g., hidden text, meta tags, or comments).
Capability inventory: The skill executes local Node.js scripts (audit-markup.mjs, observe-runtime.mjs) which have network access to crawl sites and utilize Playwright for browser automation.
Sanitization: The skill uses a custom HTML parser (scripts/lib/html.mjs) to process the DOM, providing some level of structure but not necessarily sanitizing the text against adversarial LLM instructions.
[EXTERNAL_DOWNLOADS]: The skill relies on external browser automation libraries. The scripts/lib/browser.mjs file contains logic to dynamically locate and import playwright or playwright-core from the user's environment or local project directory. If the library is missing, it provides instructions to the user on how to install it via NPM.
[DYNAMIC_EXECUTION]: The scripts/observe-runtime.mjs script utilizes page.evaluate() to execute JavaScript code within the context of the headless browser. While this is used to extract layout information (e.g., touch target sizes, horizontal overflow), it involves executing code in a browser environment based on the analyzed page's state.
[COMMAND_EXECUTION]: The skill's primary function involves the agent executing local Node.js scripts through the shell to perform audits. These scripts are provided within the skill package and manage their own logic for crawling and reporting findings.

quality-all