The Agent Skills Directory

[PROMPT_INJECTION]: The skill includes instructions in the 'Primary Directive' and 'Execution Context' sections that mandate literal interpretation and discourage human clarification (e.g., 'executed systematically without human interpretation or clarification'). These patterns are often used to override safety guidelines or bypass human-in-the-loop review.
[COMMAND_EXECUTION]: The skill provides a bash script template for identifier validation that utilizes grep, sed, sort, and uniq. It instructs the agent to define a PLAN_FILE variable based on the ${input:PlanPurpose}. If the agent executes these commands, it could be vulnerable to command injection if the input contains malicious shell characters.
[INDIRECT_PROMPT_INJECTION]: The skill is susceptible to indirect prompt injection as it processes external input to generate plans intended for autonomous execution. \n- Ingestion points: The ${input:PlanPurpose} variable is used to populate the goal and filename of the implementation plan. \n- Boundary markers: The skill lacks explicit delimiters or instructions to ignore potential commands within the user-provided purpose. \n- Capability inventory: The agent is granted the ability to write files to the /plan/ directory and execute shell-based validation commands. \n- Sanitization: There is no evidence of sanitization or validation of the ${input:PlanPurpose} string before it is used in file paths or command variables.

create-implementation-plan