The Agent Skills Directory

[PROMPT_INJECTION]: The skill contains instructions and examples related to prompt injection; however, these are used exclusively for security testing and refusal logic. The skill instructions in SKILL.md explicitly direct the agent to identify and remove injection phrases (e.g., "ignore previous instructions") from user prompts and refuse strengthening prompts designed for jailbreaking or system prompt extraction. The example in examples/refuse-extraction.md demonstrating an injection attack is correctly used to illustrate how the skill should refuse such requests.
[COMMAND_EXECUTION]: Several Python scripts used for internal automation and testing (scripts/run_agentic_request.py, scripts/sync_mise_primitives.py, and scripts/tests/test_scripts.py) utilize subprocess.run(). These calls are restricted to internal developer workflows, such as checking the version of the mise tool runner or executing unit tests for the skill's own utility scripts. They do not process untrusted user input as shell commands.
[DYNAMIC_EXECUTION]: The script scripts/run_agentic_request.py acts as a dispatcher for LLM requests, running an external runner command with specified arguments. This is part of the "Factory execution contract," a meta-instruction set for skill maintenance, and is not exposed to user-provided prompts during normal operation.
[EXTERNAL_DOWNLOADS]: The script scripts/sync_mise_primitives.py fetches JSON schema files from mise.jdx.dev and the official mise repository on GitHub. These are well-known, trusted sources associated with the mise task runner used by the skill author for project configuration management. These downloads do not contribute to a higher risk level.

prompt-enhancer