The Agent Skills Directory

[CREDENTIALS_UNSAFE]: The skill provides explicit instructions in SKILL.md to execute a shell command (cat .env | grep -E "API_KEY|_KEY") designed to retrieve and display sensitive API keys from the .env configuration file. This exposes secrets like ANTHROPIC_API_KEY, OPENAI_API_KEY, GOOGLE_API_KEY, and others to the session logs.
[EXTERNAL_DOWNLOADS]: The skill documentation instructs the user/agent to download and install the playwright package and its associated chromium browser binaries from remote sources to enable web-based automation features.
[COMMAND_EXECUTION]: The skill documentation includes shell commands for package installation and environment inspection, which an agent might execute automatically if prompted to set up or verify the environment.
[INDIRECT_PROMPT_INJECTION]: The model_router.py script serves as a gateway to multiple external AI models. It ingests user-provided content via the prompt and system_prompt parameters and interpolates them into API requests for Claude, GPT-4, Gemini, and others.
Ingestion points: The prompt and system_prompt arguments in the write and compare methods of the ModelRouter class in model_router.py.
Boundary markers: Absent. The skill does not use delimiters or instructions to prevent the target models from obeying commands embedded within the user data.
Capability inventory: Performs network operations to external API endpoints (Anthropic, OpenAI, xAI, Google, Z.AI) and executes shell commands as described in SKILL.md.
Sanitization: Absent. The inputs are passed directly to the external model clients without escaping or validation.

multi-model-writer