The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill downloads the official @playwright/cli package from the NPM registry via npx or global installation.
Evidence: scripts/playwright_cli.sh uses npx --package @playwright/cli playwright-cli to execute the tool.
Evidence: SKILL.md provides instructions for a global install via npm install -g @playwright/cli@latest.
[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted external data from websites, creating a potential surface for indirect prompt injection attacks.
Ingestion points: The agent navigates to and interacts with arbitrary URLs using pwcli open as shown in SKILL.md and references/cli.md.
Boundary markers: Absent. There are no instructions to the agent to distinguish between its internal goals and instructions potentially found on a webpage.
Capability inventory: The skill allows for high-impact actions including form filling, element interaction, executing JavaScript (eval), and writing files (screenshots, PDFs, and traces) to the output/playwright/ directory.
Sanitization: Absent. Web content is rendered and interacted with directly by the browser tool.
[DYNAMIC_EXECUTION]: The skill exposes commands that allow for the execution of arbitrary JavaScript code within the context of the automated browser.
Evidence: references/cli.md documents pwcli eval (for running JS snippets) and pwcli run-code (for executing Playwright-specific code blocks).
[COMMAND_EXECUTION]: The skill utilizes a shell script to wrap and execute the Playwright CLI with arguments provided by the agent.
Evidence: scripts/playwright_cli.sh constructs a command array and uses exec "${cmd[@]}" to run it.

openai-playwright