Skill Optimizer: STOM Automation Engine

Problem

Many agent skill manifests grow beyond what is needed to trigger a task. As SKILL.md files accumulate examples, API notes, edge-case rules, and long checklists, they can burn tokens and become harder for weaker models to follow. Manual review is easy for a single file but brittle when applied across a skill library.

How It Works

• Single-skill audit: reads a target SKILL.md and scores it against STOM v2.0 dimensions, including size, information density, layering, architecture health, and model adaptation.
• Batch audit: scans multiple skills, ranks them by severity, and surfaces the files most likely to benefit from compression or restructuring.
• Controlled optimization: audit workflows are read-only; only the optimize workflow modifies files. Changes focus on moving, compressing, or externalizing content while preserving business logic.
• Weak-model adaptation: v1.2.0 checks M1-M10 deviation patterns, using static rules for ambiguity and step clarity, plus LLM review for instruction interference, thought leakage, and convention-vs-novelty issues.

Boundaries

This skill targets information architecture and token efficiency in SKILL.md, not functional correctness or new business capabilities. It is useful for auditing existing skills, cleaning up oversized entries, and planning new skill structure, but it should not be used to rewrite code behavior or invent missing product logic.