| name | semantic-organization |
| description | Governs how every skill in this marketplace is structured. Codifies the Anthropic Agent Skills directory convention (skills/<name>/SKILL.md at the repo root — no plugin wrapper), folder roles, naming rules, and migration triggers. Scaffolds new skills against the canonical layout, audits existing skills, and recommends when a folder should become its own sibling skill. Applies Unix philosophy: one skill = one purpose, compose via plain text, prefer small over large. Activates on "scaffold a new skill", "/new-skill", "audit this skill", "/semantic-audit", "should this be a folder or its own skill", or any time a skill structure is being designed or evaluated.
|
semantic-organization
Purpose
Every skill in this marketplace must match the Anthropic Agent Skills
directory convention at https://github.com/anthropics/skills. This skill
encodes that convention and audits against it.
Canonical layout
<marketplace-repo>/
├── .claude-plugin/
│ └── marketplace.json ← marketplace manifest (lists skills)
├── skills/ ← all skills live here (mirrors anthropics/skills)
│ └── <skill-name>/
│ ├── .claude-plugin/
│ │ └── plugin.json ← plugin manifest (inside the skill)
│ ├── SKILL.md ← required by spec
│ ├── README.md ← marketplace listing (≤200 words)
│ ├── TESTS.md ← end conditions + test cases
│ ├── scripts/ ← spec-canonical optional folder
│ ├── references/ ← spec-canonical optional folder
│ ├── assets/ ← spec-canonical optional folder
│ ├── prompts/ ← convention (mode-specific operations)
│ ├── templates/ ← convention (output shapes)
│ ├── schemas/ ← convention (JSON Schema contracts)
│ └── fixtures/ ← convention (test inputs + expected outputs)
├── README.md
└── LICENSE
Forbidden layouts
These are NOT canonical. Audit returns broken if found:
| Forbidden | Reason |
|---|
plugins/<name>/skills/<name>/SKILL.md | Old wrapped form; double nesting |
plugins/<name>/SKILL.md | Wrong directory name; should be skills/ |
<repo-root>/SKILL.md (skill at repo root) | Multi-skill repos must use skills/<name>/ |
The two layers — same directory
A skill carries both layers in the same folder:
| Layer | Files | Source |
|---|
| Spec layer (Anthropic) | SKILL.md, optional scripts/, references/, assets/ | https://agentskills.io/specification |
| Marketplace layer (Claude Code) | .claude-plugin/plugin.json, README.md, TESTS.md | This skill's conventions |
There is no separate plugin-wrapper directory. The plugin manifest lives
INSIDE the skill folder, so the skill is simultaneously
spec-compliant AND installable via claude plugin install.
Skill profiles
Not every skill needs every folder. Two profiles:
Single-rule skill
One durable rule or principle. SKILL.md body carries everything.
Required: SKILL.md, .claude-plugin/plugin.json, README.md, TESTS.md.
No required sub-folders.
A skill is single-rule when SKILL.md is under 200 lines and there are no
sub-folders.
Full-shape skill
Multiple modes, references, optional scripts and assets. SKILL.md is the
entry point; detail lives in references/ and assets/.
Required: same as single-rule, plus at least one of references/,
scripts/, assets/ per actual need.
A skill is full-shape when it has multiple operating modes, references
broken out from SKILL.md, and (usually) prompts or assets.
The eight audit dimensions
See references/audit-rubric.md.
| Layer | Dimension |
|---|
| Spec | S1. SKILL.md presence and validity |
| Spec | S2. Naming conformance (parent dir, plugin.json, frontmatter agree) |
| Spec | S3. Body discipline (<500 lines; shallow references) |
| Spec | S4. Folder discipline (only spec or convention folders; no forbidden) |
| Marketplace | P1. Plugin manifest valid (inside skill folder) |
| Marketplace | P2. README discipline (≤200 words, what/when/install) |
| Marketplace | P3. TESTS.md presence and quality |
| Marketplace | P4. Migration health (no overgrown folders; no wrapper layers) |
Verdicts
| Verdict | Rule | Allowed downstream |
|---|
spec-compliant + marketplace-ready | All 8 dims ≥4 | Ship |
spec-compliant, marketplace-drift | Spec ≥4, marketplace has 2-3 | Fix conventions before ship |
spec-drift | One or more spec dims at 2-3 | Fix spec violations first |
broken | Any dim at 1, OR a forbidden layout is present | Halt; restore canonical shape |
Operating modes
Scaffold mode
Trigger: "scaffold a new skill", "/new-skill", "create a skill"
Behavior:
- Ask for skill name (kebab-case, matches spec rules)
- Ask for one-sentence description (what + when)
- Ask for profile: single-rule or full-shape
- Emit canonical folder tree at
skills/<name>/ with placeholder content
- Run immediate self-audit — confirm scaffold scores ≥4 on all dimensions
See prompts/scaffold-new-skill.md.
Audit mode
Trigger: "audit this skill", "/semantic-audit"
Behavior:
- Walk the skill's file tree at
skills/<name>/
- Detect profile (single-rule vs. full-shape)
- Check for forbidden layouts (
plugins/ wrapper, double nesting)
- Score each of the 8 dimensions per
references/audit-rubric.md
- Emit Semantic Organization Audit with findings
- Verdict per the table above
See prompts/audit-existing-skill.md.
Migration mode
Trigger: "should this be a folder or its own skill?"
Behavior:
- Check folder against migration triggers
- Return:
stay-as-folder / promote-to-sibling-skill / already-its-own-skill
- If promote, emit migration plan
See references/migration-triggers.md and prompts/evaluate-migration.md.
Rename mode
Trigger: "rename this folder", "is this folder name semantic?"
Behavior:
- Compare folder name against the role taxonomy
- Propose a spec-canonical name or flag for deletion
- Refuse to rename spec-canonical folders (
scripts/, references/, assets/) to non-spec names
See prompts/rename-for-semantics.md.
References
references/spec-vs-conventions.md — how spec and marketplace layers coexist in one directory
references/folder-roles.md — what each canonical folder holds; forbidden names
references/naming-rules.md — spec-mandated and convention-mandated naming
references/migration-triggers.md — when a folder becomes its own skill
references/audit-rubric.md — the 8-dimension scored rubric
references/unix-philosophy.md — Unix tenets applied to skills
Triggers
| Phrase | Mode |
|---|
scaffold a new skill, /new-skill, create a skill | Scaffold |
audit this skill, /semantic-audit | Audit |
should this be a folder or its own skill? | Migration |
rename this folder, is this folder name semantic? | Rename |
Self-application
This skill must pass its own audit. The audit applies the 8-dimension
rubric. Self-test target: all skill-layer dimensions at 5; all
marketplace-layer dimensions at 5; verdict =
spec-compliant + marketplace-ready.
If the skill cannot meet its own bar, fix the rubric or fix the skill.