Recipe: Modularization Audit for plan-marshall Skills

Generic, domain-invariant recipe skill that drives a plan to audit a corpus of skill documents for modularization opportunities — over-long inline step sequences that belong in referenced sub-documents, numbering defects, and document bloat. Where recipe-marshal-json-config-audit audits the live .plan/marshal.json config file and recipe-plugin-compliance sweeps marketplace bundles for architecture compliance, this recipe inspects every SKILL.md and sub-document beneath a resolved target root and collects one remediation deliverable per offending skill.

Like recipe-marshal-json-config-audit it is an LLM-driven, SKILL.md-only deliverable-collection workflow (no scripts): the phase-3-outline recipe path loads it to produce a modularization-audit solution_outline.md.

This recipe has two orthogonal, independent dials at its runtime:

• A hard-coded COVERAGE CELL (thoroughness=T5, scope=overall) — the depth and breadth of the audit. It is NOT gathered from the user. The recipe implements the coverage-gathering contract — expand and consume — but skips the gather step for the cell, supplying the fixed identifier + expanded instruction per the contract's gather → expand → consume model.
• A gathered CORPUS ROOT — the tree the audit is rooted at. It IS gathered, via AskUserQuestion (Step 1). The cell governs HOW DEEPLY/BROADLY the audit runs; the corpus root governs WHAT TREE it runs over. The two are distinct and never conflated: the no-gather prohibition applies to the CELL ONLY, never to the target.

Both the resolved cell and the resolved target root are persisted to status.json metadata.

Input Parameters