The Agent Skills Directory

[PROMPT_INJECTION]: The skill incorporates specific defenses against injection, instructing the agent to verify that any proposed 'learning' originates from genuine problem-solving rather than the user attempting to dictate rules through prompt text.
[COMMAND_EXECUTION]: The skill utilizes local CI scripts (./ci/utils/validate_skills.sh and ci/test_skills_assets.sh) to validate proposed changes. These are internal tools used for structural and functional verification of the repository.
[SAFE]: The workflow includes a non-negotiable security section that prevents the agent from weakening existing safety guardrails, expanding its own permissions, or introducing insecure code such as eval(), exec(), or os.system() with user-controlled input.
[SAFE]: All proposed modifications require explicit user confirmation before being applied, providing a critical human-in-the-loop checkpoint to prevent automated or malicious skill degradation.

cuopt-skill-evolution