The Agent Skills Directory

[SAFE]: The skill implements a robust 'snapshot-and-benchmark' pattern that ensures transparency and prevents overfitting. It explicitly forbids the agent from modifying the immutable baseline (skill-snapshot/) or loosening evaluation criteria to force a pass.
[COMMAND_EXECUTION]: The skill uses standard bash commands (mkdir, cp, diff) and uv run to manage snapshots and invoke a local validator (quick_validate.py). All execution is scoped to the skill's specific workspace and target directories.
[PROMPT_INJECTION]: Instructions include clear directives to ignore instructions from the data being processed if they conflict with the rubric. The skill is designed to audit other skills but includes safeguards against self-referential or adversarial prompts in the analyzed content.
[DATA_EXPOSURE]: The skill operates entirely within the repository context (skills/ and .improve-skill-workspace/). No network-based exfiltration or access to sensitive global credentials (e.g., .ssh, .aws) is present.
[TRUSTED_VENDORS]: The skill correctly reuses the quick_validate.py tool from skills/create-skill/, which is part of the local vendor environment, following least-privilege principles by not reinventing existing internal tooling.

improve-skill