The Agent Skills Directory

[DYNAMIC_EXECUTION]: The skill is designed to generate, register, and execute arbitrary code to solve unique problems.
Evidence: Steps 1, 2, and 4 describe identifying needs, writing Python scripts (e.g., tools/find_x.py), implementing Node.js MCP servers using the @modelcontextprotocol/sdk, and registering them in the project_tools configuration for active use.
[INDIRECT_PROMPT_INJECTION]: Tools constructed by this skill are intended to process external data, creating a vulnerability surface for indirect prompt injection.
Ingestion points: The skill explicitly mentions tools for migrating CSV data, querying internal Knowledge Bases, parsing logs, and interacting with external APIs like JIRA (SKILL.md).
Boundary markers: Instructions for implementing boundary markers or "ignore instructions" delimiters are absent from the tool generation guidelines.
Capability inventory: Generated tools are granted capabilities for file system modification, network API requests, and persistent registration within the agent's startup context (SKILL.md).
Sanitization: The skill advises the agent to "rigidly enforce static analysis" and "strictly prohibit the use of arbitrary eval functions," though specific data sanitization logic for ingested content is not detailed.
[COMMAND_EXECUTION]: The skill contains instructions regarding the risk of command injection in dynamically created tools.
Evidence: The "Skill Security" section explicitly warns against generating tools that use shell=True in Python or unsanitized exec() calls in JavaScript, identifying them as potential Remote Code Execution (RCE) backdoors.

tool-construction