Skill Optimizer

Problem

SKILL.md files often become too long, trigger ambiguously, contain contradictory instructions, or inline templates and checklists that consume context unnecessarily. The agent may under-trigger, over-trigger, or still need follow-up questions, while token usage remains high. skill-optimizer treats this as an engineering cleanup task for an existing Skill, not as a generic prompt rewrite.

How It Works

It follows a structured optimization path:
- Diagnosis: scores five dimensions: structure, SKILL.md length, description quality, instruction redundancy, and design-pattern fit.
- Structure: splits content into metadata, SKILL.md, and resource layers; moves long checklists, examples, templates, and error-handling lists to references/ or assets/.
- Pattern selection: applies tool wrapper, generator, reviewer, inverter, or pipeline patterns, including combinations.
- Instruction pruning: removes repetition, contradictions, and over-constrained rules; prefers explaining why and using examples instead of stacked directives.
- Trigger tuning: rewrites description to clarify capability boundaries, keywords, and use cases, reducing under-triggering.

Boundaries

It is best suited to Agent/CodeBuddy-style Skills where a SKILL.md, file path, or pasted content is available. Without source material it can only request clarification. If the root issue is model capability, tool permissions, or business logic, it focuses on prompt structure and may not replace implementation fixes.