The Agent Skills Directory

PROMPT_INJECTION (MEDIUM): The files persuasion-principles.md and CLAUDE_MD_TESTING.md contain explicit instructions and frameworks designed to manipulate AI behavior and override system defaults.
persuasion-principles.md serves as a guide for using 'Authority', 'Scarcity', and 'Commitment' to bypass AI hesitation and enforce compliance ('YOU MUST', 'No exceptions'). It explicitly references research on persuading AI to comply with 'objectionable requests'.
CLAUDE_MD_TESTING.md uses high-pressure scenarios (e.g., 'Production is bleeding money') and absolute failure conditions ('If a skill existed... and you didn't use it, you failed') to force specific agent behaviors.
COMMAND_EXECUTION (MEDIUM): The script render-graphs.js utilizes child_process.execSync to execute the system dot command (Graphviz). While it passes the graph content via the input option (reducing shell injection risk), it still allows for the execution of arbitrary Graphviz logic and relies on system binaries.
INDIRECT_PROMPT_INJECTION (LOW): The csv-data-analyzer-example.md documentation describes a workflow that proactively processes external .csv files and generates reports. This creates an attack surface where instructions embedded in user-uploaded data could potentially influence the agent's next steps, although the provided template includes basic validation tags.
DYNAMIC_EXECUTION (LOW): The skill example csv-data-analyzer-example.md references the execution of a Python script (analyze_csv.py) that is not provided within the skill files, implying a dependency on external or dynamically available scripts for core functionality.

writing-skills