The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill is designed to ingest tool specifications and failure examples to generate optimized tool descriptions and error messages, which creates a potential surface for indirect injection if processing untrusted specs.
Ingestion points: The ToolSchemaBuilder in scripts/description_generator.py and the optimize_tool_description prompt template in SKILL.md accept external metadata.
Boundary markers: The skill uses standard markdown headers and triple-backtick blocks as delimiters in generated content.
Capability inventory: The skill provides guidance for building high-permission tools, such as the execute_command bash tool mentioned in the architectural reduction case study.
Sanitization: The ErrorMessageGenerator correctly uses json.dumps to ensure structured outputs are valid JSON, though the skill relies on the underlying LLM's reasoning for sanitizing natural language descriptions.
[COMMAND_EXECUTION]: The skill advocates for the "File System Agent Pattern," which involves providing an agent with bash command execution capabilities (e.g., ls, grep, cat) to navigate documentation and code.
The instructions explicitly emphasize that this pattern should be implemented within a restricted sandbox environment to mitigate security risks.
The provided Python scripts do not perform any command execution; they are limited to metadata processing and string generation tasks.

tool-design