The Agent Skills Directory

[COMMAND_EXECUTION]: The scripts/test-hook.sh utility executes local scripts provided as command-line arguments. While this is the intended behavior for a developer testing tool, it facilitates the execution of arbitrary local shell scripts.
[PROMPT_INJECTION]: The skill recommends a 'prompt hook' pattern that interpolates untrusted tool inputs (such as shell commands or file paths) directly into prompts for LLM evaluation without sanitization or boundary markers.
Ingestion points: Untrusted data from tool inputs is interpolated into prompts as described in SKILL.md, references/migration.md, and references/patterns.md (e.g., using the $TOOL_INPUT.command variable).
Boundary markers: Provided examples do not implement delimiters, XML tags, or specific instructions to the LLM to ignore embedded commands within the untrusted input.
Capability inventory: Decisions made by these prompts determine whether the agent allows or blocks subsequent operations, including file system modifications and system command execution.
Sanitization: There is no evidence of validation, escaping, or filtering applied to the untrusted data before it is interpolated into the prompt context.

hook-development