The Agent Skills Directory

[COMMAND_EXECUTION]: The scripts/prompt_tester.py script implements a feature to call external LLM runners using the --runner-cmd argument. The implementation mitigates injection risks by formatting the command string and then processing it through shlex.split() before executing via subprocess.run() with shell=False. This design ensures that dynamic components like prompts and test inputs are treated as literal arguments rather than shell instructions.\n- [INDIRECT_PROMPT_INJECTION]: The skill processes external data (prompts and test cases) to evaluate AI performance, creating a surface for indirect prompt injection if the source data is untrusted.\n
Ingestion points: testcases.json (via --cases-file), user-provided prompt files, and stdin JSON payloads.\n
Boundary markers: The skill does not currently implement specific boundary markers or instructions to ignore embedded commands within the processed inputs.\n
Capability inventory: The toolkit can invoke arbitrary external CLI tools and write prompt history to local files.\n
Sanitization: No input sanitization is applied, as the tool is designed to process arbitrary text for evaluation purposes.\n- [PERSISTENCE]: The scripts/prompt_versioner.py script uses a local file .prompt_versions.jsonl to store a history of prompt revisions. This is a standard and safe implementation for local configuration and history tracking.

prompt-engineer-toolkit