The Agent Skills Directory

[SAFE]: The skill implements a legitimate prompt optimization workflow using reputable libraries like agentlightning and opto (Microsoft). All operations are focused on local prompt engineering tasks.
[INDIRECT_PROMPT_INJECTION]: The skill ingests external datasets (JSONL) and interpolates them into prompt templates. This surface is mitigated by strict placeholder validation in scripts/optimize_support.py (e.g., validate_template_against_task) and brace-escaping logic in render_prompt_text to prevent template injection.
[COMMAND_EXECUTION]: Command execution is limited to internal Python scripts (run_optimize.py, train.py) and standard library usage for local development. Subprocess logic is handled through the github-copilot-sdk using the current logged-in user session, adhering to standard local security practices.
[DATA_EXPOSURE]: The skill reads .env files for configuration (specifically COPILOT_MODEL). This is standard behavior for project-level tooling and does not involve exfiltrating sensitive environment variables to external domains.

trainer-optimize