The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill instructs the agent to process data from external, potentially untrusted sources which could be used to influence the agent's behavior.
Ingestion points: The agent reads game state information via window.render_game_to_text, browser console errors, and a local progress.md file that may contain instructions from previous sessions or actors.
Boundary markers: The instructions lack robust delimiters or specific directives to ignore embedded instructions within these data sources, though they suggest identifying an "Original prompt" prefix in the progress log.
Capability inventory: The agent is granted access to high-impact tools such as Bash, Write, and Edit, enabling it to modify the file system or execute shell commands based on the data it ingests.
Sanitization: No explicit sanitization, validation, or escaping of the ingested content is mandated before the agent evaluates it.
[EXTERNAL_DOWNLOADS]: The skill provides instructions for the agent to install external software to ensure the testing environment is properly configured.
Evidence: The workflow includes a fallback instruction to install the @playwright/mcp package globally using npm install -g @playwright/mcp@latest.
[COMMAND_EXECUTION]: The workflow requires the agent to execute a local Node.js script to automate browser tasks and capture output.
Evidence: The agent must run the provided $WEB_GAME_CLIENT script (located at scripts/web_game_playwright_client.js) to interact with the game and generate test artifacts.

openai-develop-web-game