The Agent Skills Directory

[COMMAND_EXECUTION]: The skill instructs the agent to execute standard development commands like pnpm run validate and pnpm run eval. These are used locally to ensure that skill files adhere to required schemas and pass behavioral test cases.
[EXTERNAL_DOWNLOADS]: The skill references established research and documentation from well-known technology entities, such as Microsoft Research (SkillOpt), Google (Chrome Eval Design), and the author's own repository (arenukvern/skill_steward). These external links are used to provide the agent with architectural guidance and best practices.
[DATA_INGESTION_SURFACE]: The skill's primary function involves analyzing and modifying other SKILL.md files, which constitutes a surface for processing potentially untrusted data.
Ingestion points: The skill reads content from the skills/** directory for evaluation.
Boundary markers: The instructions mandate a 'manual adaptation' loop and 'bounded edits' (restricting changes to ≤10% line churn), which facilitates human oversight and reduces the risk of automated obedience to embedded instructions.
Capability inventory: The skill uses local shell commands (pnpm, plugin-eval) for validation and benchmarking.
Sanitization: The skill emphasizes structural validation (Layer 0) before any complex behavioral analysis.

skill-eval-improve