Skill Sifter

Problem It Solves

SKILL.md files can look complete while still being hard for an agent to execute reliably. The document may mix purpose, instructions, examples, schemas, and error handling without a clear sense of what is redundant, what is missing, and whether the frontmatter defines the calling boundary well enough. This skill turns subjective review into checkable questions: whether INSTRUCTION content is sufficient, whether paths, variables, schemas, and ID strategies are consistent, and whether error handling is actionable.

How It Works

It evaluates a skill in four progressive layers. Positioning first reads user-invocable, allowed-tools, dependencies, and intended scenarios to build a one-line reference model. Structure audit then classifies the document into META, PURPOSE, INSTRUCTION, SCHEMA, EXAMPLE, CODE, and ERROR sections, checking balance and information density. Defect detection scans consistency, necessity, and fitness across language, granularity, and scenario coverage. Optimization finally ranks actionable changes, such as adding an agent operation manual, layering schemas, externalizing code, improving error handling, and unifying conventions.

Boundaries

It is useful for reviewing new skills, diagnosing weak skills, comparing options, and setting release gates. Its confidence depends on the material available; if a skill relies on external services without documented interfaces, the assessment is limited. For multi-document systems, it can inspect references/ and scripts/, but SKILL.md self-containment remains the primary criterion. It does not replace runtime tests, and scores should be calibrated against tool permissions, model behavior, and production feedback.