Skill Compliance Enforcer

When your agent acknowledges your skill file but does whatever it wants anyway, this auditor diagnoses the structural failure and tells you exactly what to fix — before you publish.

You wrote a SKILL.md with careful instructions. The agent loaded it. You can see it read the file. Then it proceeded to ignore every constraint, write memory entries instead of following the skill, and burn tokens on rework.

This isn't a content quality problem — it's structural. Real GitHub issues document 100% failure rates (issue #91391: 5 worktrees × 10 repeats, 20M tokens lost). Issue #46765 aggregates 9 independent failure reports from users with 500+ hours experience. The root causes are consistent: constraints buried too deep in the file, missing blocking directives, no pre-commitment checkpoints, and agent-memory conflicts that silently override skill content.

• 20-Check Structural Audit — Categorizes failures into Attention Window, Constraint Enforcement, Structural Integrity, and Trigger/Invocation issues

• Compliance Score out of 100 — Quantified assessment with 5 severity tiers from "publication-ready" to "major restructure"

• Priority Fix Order — Ranked by actual compliance impact, not alphabetical or arbitrary

• Concrete Fix Recommendations — Every FAIL gets a specific, actionable fix (not vague advice like "make it clearer")

• Anti-Memory-Conflict Detection — Catches the silent killer where agent memory overrides skill instructions

• Constraint Placement Analysis — Maps where your blocking directives appear relative to the attention window

Asking Claude "review my SKILL.md for quality" returns "looks great!" regardless of actual compliance. The agent can't diagnose its own attention-window behavior. Telling it "be strict" still misses structural issues because the model doesn't have visibility into its own processing priorities.

This skill encodes 20 structural patterns derived from 20+ documented failure cases. It checks constraint