The Agent Skills Directory

[COMMAND_EXECUTION]: The skill utilizes local Python scripts (scripts/ab_harness.py, scripts/preference_log.py, and scripts/judge_train_stub.py) to manage its A/B testing workflow and preference logging. These scripts perform deterministic operations such as shuffling variant labels, rendering side-by-side comparisons to the terminal, and recording user choices in a JSONL file. Analysis shows these scripts use standard libraries and do not involve unsafe shell execution or remote code loading.
[SAFE]: The skill features a robust 'Constitutional Judging' phase (Phase 2) that acts as a security and logic floor. It automatically evaluates generated variants against the target skill's 'constitution' (MUST/MUST NOT clauses) and rejects any that violate defined behavioral guarantees before they reach the user for evaluation.
[SAFE]: No network activity, external dependencies, or remote resource fetching were detected. The skill maintains data locality by storing all logs and variant outputs within the local workspace environment.

skill-tuning