The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill documents how to process tool inputs and session transcripts through prompt-based hooks. This establishes a surface for indirect prompt injection if user-provided or external data contains adversarial instructions.\n
Ingestion points: Hook scripts ingest $TOOL_INPUT and $TRANSCRIPT_PATH via stdin or environment variables.\n
Boundary markers: Documentation examples show interpolation of input data directly into prompts; users are encouraged to apply reasoning to validate these inputs.\n
Capability inventory: The skill includes examples for file system access and shell execution, intended for validation and automation.\n
Sanitization: The skill provides a hook-linter.sh utility that checks for variable quoting and other shell safety best practices to prevent injection in command hooks.\n- [COMMAND_EXECUTION]: The utility scripts/test-hook.sh is designed to execute user-provided scripts for testing and debugging purposes. This is the primary function of the script and is intended for local developer use.\n- [DATA_EXFILTRATION]: Documentation in references/advanced.md provides examples of integrating hooks with external services, such as Slack webhooks and remote databases for logging and auditing. These patterns illustrate legitimate integration capabilities for automated workflows.

hook-development