The Agent Skills Directory

[COMMAND_EXECUTION]: The skill's core functionality involves executing shell commands to verify project state and behavior.
Evidence: The command-check verification method executes arbitrary strings via Bash (e.g., npm run typecheck).
Evidence: The behavior-check method executes shell scripts (e.g., bash bench/run-contract-case.sh) to verify complex assertions.
[DYNAMIC_EXECUTION]: The skill allows for the execution of dynamically defined or referenced scripts.
Evidence: The script-check method executes mcp-scripts defined in project configuration files, which can contain arbitrary JavaScript or other logic.
Evidence: The skill-check method invokes a Python script at a specific path: .claude/skills/skill-creator/scripts/quick_validate.py.
[INDIRECT_PROMPT_INJECTION]: The skill is vulnerable to indirect injection if an attacker can modify the files it processes to trigger malicious commands.
Ingestion points: The skill reads and iterates through all files in the .evals/ directory and the EVAL_INDEX.md file.
Boundary markers: There are no explicit delimiters or sanitization routines mentioned for the content of these eval cases before they are used to construct shell commands.
Capability inventory: The skill has the ability to run bash commands, execute scripts, and read/write to the local filesystem.
Sanitization: No sanitization is performed on the command, pattern, or target fields extracted from the markdown eval cases.

eval-creator