The Agent Skills Directory

[COMMAND_EXECUTION]: The skill leverages the gbrain CLI utility to manage the optimization lifecycle, including bootstrapping benchmarks and performing validation-gated mutations on skill instruction files. These commands are integral to the skill's stated purpose of self-optimization.
[INDIRECT_PROMPT_INJECTION]: The skill processes instructions from other SKILL.md files and benchmark datasets. This represents an attack surface where malicious input in a processed skill could theoretically influence the optimizer's output. To mitigate this, the skill implements explicit safeguards: a mandatory human review step (enforced via a sentinel string) for all bootstrapped content and strict validation gating (median-of-3 performance checks) before any instructions are updated.
[DATA_EXPOSURE]: The skill records its operations in audit logs located at ~/.gbrain/audit/ and maintains a history of versioned edits and rejected candidates within the skill's local directory. These operations are localized and part of the skill's transparency mechanism.

skill-optimizer