The Agent Skills Directory

[PROMPT_INJECTION]: The document references/failure-modes.md contains common prompt injection patterns such as "ignore previous instructions". This text is part of a catalog of failure modes and is used for educational and defensive design purposes rather than as an attempt to hijack the agent.\n- [COMMAND_EXECUTION]: The scripts assets/contract_check.py and eval/run_eval.py use the Python subprocess module to execute system commands. This is primarily used for environment validation and interacting with Git for contract and grant verification.\n- [DYNAMIC_EXECUTION]: The contract_check.py script is designed to execute shell commands extracted from the assertions field of JSON 'envelope' files. While this creates a dynamic execution surface, the skill instructions specifically direct the agent to list these commands and wait for user confirmation before proceeding with execution.

crafting-self-prompting-loops