claude-plugin-structure-check - SKILL.md Agent Skill

name: claude-plugin:structure-check description: >- Audit a claude-plugin marketplace repo (e.g. claude-plugin-visuals) against the standard directory layout and report PASS/WARN/FAIL/N/A. Supports both mono (plugins/<p>/skills/) and single (repo-root skills/) layouts — auto-detected, or forced with --single / --mono. Read-only — never edits. Discovers plugins/skills dynamically by directory scan, then evaluates mandatory items M1-M6 (FAIL) and recommended items R1-R5 (WARN). Use when the user says "check my claude-plugin repo structure", "is this marketplace repo standard?", "audit plugin layout", "/claude-plugin:structure-check". Sister skills: claude-plugin:structure-refactor (fixes what this finds), claude-plugin:rename-repo (renames the repo to the team convention). Do NOT use for SKILL.md content quality (use skill:check) or shell scripts (use sh:check). compatibility: tools: Read, Glob, Grep, Bash metadata: model_recommendation: tier: haiku reason: "read-only directory-structure audit; dynamic scan + JSON/frontmatter validation; bounded PASS/WARN/FAIL report" claude: prefer non_claude: advisory-only

claude-plugin Structure Auditor

Help

If arg #1 is -h, --help, or help, read references/help.md and output its content verbatim, then stop. No filesystem scan.

Step 1: Parse Args + Resolve Repo Path

Positional [repo-path] → audit that path. None → audit the current directory.
Flags --single / --mono → force the layout mode, overriding auto-detection (Step 2). Mutually exclusive; if both given, last wins.
Confirm the path exists; if not, stop with a one-line error pointing to /claude-plugin:structure-check path/to/repo.
Check test -d <path>/.git; not a git repo → continue, but note it (one warning line — the audit still runs).

Step 2: Detect Mode + Discover Plugin Roots + Skills

Read references/structure-spec.md for the full standard (embedded SSOT) — see "Layout modes", "Mode detection", and "Mandatory items by mode".

For detailed evaluation rules and mode/type classification logic: see references/evaluation-rules.md

Record the detected mode, plugin-root list, and skill list for the report header and the per-skill recommended checks (R1/R2/R5).

Step 3: Evaluate M1-M6 and R1-R5

Apply PASS/WARN/FAIL/N/A to each item per the scoring rules in references/evaluation-rules.md.

Step 4: Output the Report

Read references/report-template.md for the exact format. The report has a header line (path + detected mode + discovered plugins/skills), a [필수] block (M1-M6) and a [권장] block (R1-R5), then the summary verdict:

any FAIL → FAIL
no FAIL but ≥1 WARN → WARN
all PASS/N/A → PASS

Emit the next-action hint only when there is ≥1 FAIL or WARN.

Constraints

Read-only — never create, move, or edit any file. Fixing is claude-plugin:structure-refactor's job.
N/A is not FAIL — a missing subject (no skills, no plugins beyond M2) yields N/A for dependent checks.
Do not audit SKILL.md content (that is skill:check) or shell script quality (that is sh:check) — only the directory structure.
Repo-agnostic: detect mode + discover plugin roots/skills by scan; the spec is embedded, not read from the target repo. A --single/--mono override means "score by that mode" — a wrong override surfaces as a normal M2 FAIL, never a silent skip.