The Agent Skills Directory

[PROMPT_INJECTION]: The skill uses a 'First-action contract' and a 'Receipt' verification system. These instructions are designed to force the AI agent to execute a specific CLI tool and quote a machine-generated stderr line before responding, which overrides the model's standard conversational flow.
[COMMAND_EXECUTION]: The skill requires the execution of a Node.js script (index.js) as the primary mechanism for detecting browser states and providing setup guidance.
[DYNAMIC_EXECUTION]: The implementation utilizes Playwright's page.evaluate and page.route methods to execute JavaScript and intercept network traffic within the user's active browser sessions for the purpose of state classification.
[INDIRECT_PROMPT_INJECTION]: The skill has an attack surface for indirect prompt injection as it ingests and processes text and HTML from external web-AI provider pages to determine session status.
Ingestion points: Browser tab content (titles, innerText, and HTML fragments) scraped via the snapshotPage function in index.js.
Boundary markers: None identified; the scraped content is processed directly against classification regexes.
Capability inventory: Execution of Node.js scripts and browser automation via Playwright.
Sanitization: Relies on specific regex pattern matching (e.g., searching for 'Access Verification' or 'organization has been disabled') to classify states.

agentweb-setup