The Agent Skills Directory

[PROMPT_INJECTION]: The skill is susceptible to indirect prompt injection because its primary function is to ingest untrusted data from external websites.
Ingestion points: The page-collect-bundle.js script extracts document.title, metadata tags, and visible body text from arbitrary URLs.
Boundary markers: Extracted data is written to structured JSON files (metadata.json, text.json), but the individual string values within these files are raw and do not contain internal delimiters or 'ignore' instructions.
Capability inventory: The skill can execute shell commands via playwright-cli and write files to the local filesystem using node:fs (found in scripts/page-collect.js).
Sanitization: The icon collector performs regex-based optimization on SVG content, but general text and metadata extraction does not perform sanitization of the scraped content. The skill provides an 'External content warning' in the documentation.
[COMMAND_EXECUTION]: The skill executes the playwright-cli tool to interact with the browser and perform extraction.
Evidence: spawnSync('playwright-cli', ...) is used in scripts/page-collect.js to open target URLs and run extraction logic. The skill supports both http(s):// and file:// protocols, the latter allowing the tool to be used for auditing local files.
[EXTERNAL_DOWNLOADS]: The browser-side bundle performs network requests to resolve remote assets.
Evidence: The resolveImgSvg function in scripts/page-collect-bundle.js uses the browser's fetch() API to download SVG image data from external sources during the icon collection process.

page-collect