The Agent Skills Directory

[COMMAND_EXECUTION]: The skill advocates for an 'Architectural Reduction' pattern where specialized tools are replaced by 'direct file system access through a single command execution tool.'
Evidence: The text explicitly recommends providing agents with standard Unix utilities (grep, cat, find, ls) to operate on the system.
Risk: This pattern significantly increases the agent's capability to perform unauthorized actions or exfiltrate data if the command execution environment is not strictly sandboxed.
[PROMPT_INJECTION]: The skill describes an 'Indirect Prompt Injection' surface through two proposed patterns.
Ingestion points: The 'File System Agent Pattern' (reading arbitrary system files) and the 'Tool-Testing Agent Pattern' (analyzing 'failure_examples' which may contain untrusted data from logs or API errors).
Boundary markers: The documentation mentions safety constraints but does not provide specific implementation details for boundary markers or instruction isolation.
Capability inventory: The design pattern encourages the use of shell execution tools and agent-driven optimization of its own prompts.
Sanitization: No specific sanitization or escaping techniques are detailed for the proposed Python code snippet optimize_tool_description, which interpolates failure data directly into an agent prompt.

tool-design