Skill creator

Skill Creator

Create new skills with proper structure, compliance, and registration.

Path convention: Commands below assume cwd is $CLAUDE_PROJECT_DIR (project root). Prefix paths with "$CLAUDE_PROJECT_DIR/" when invoking from subdirectories.

When to Use

• User asks to "create a skill", "build a new skill", "make a skill for [X]"
• Converting external skill for adoption
• Scaffolding a skill from a workflow pattern

Explicit: /mk:skill-creator [name] [description]

Workflow Integration

Meta skill — not tied to a specific Phase. Invoked on-demand for skill development.

Scripts

Script-first approach — Python scripts handle scaffolding and validation. Claude reviews and fills in the content.

# Scaffold a new skill (creates directory + template SKILL.md)
.claude/skills/.venv/bin/python3 .claude/skills/skill-creator/scripts/init-skill.py mk:my-feature --path .claude/skills

# Validate an existing skill against compliance checklist
.claude/skills/.venv/bin/python3 .claude/skills/skill-creator/scripts/validate-skill.py .claude/skills/my-feature

Process

1. Gather intent — what should the skill do? When should it trigger? What output format?
2. Scaffold — run init-skill.py to create directory + template SKILL.md with TODO markers
3. Fill content — Claude completes each TODO section in the generated template
4. Add references/ — if skill body would exceed ~100 lines, split into references
5. Security boundaries — load mk:skill-template-secure template for trust model if needed
6. Validate — run validate-skill.py to check compliance (8-point checklist)
7. Fix failures — if validation score < 7/8, fix failing items
8. Register — add row to SKILLS_ATTRIBUTION.md
9. Report — output creation summary

Required Sections in Generated SKILL.md

Every skill MUST have:

# Frontmatter
name: "{name}"
description: "{specific trigger keywords + what it does}"

If the name already exist then use with prefix mk:

# Frontmatter
name: mk:{name}
description: "{specific trigger keywords + what it does}"

## Overview — what + when (2-3 sentences)

## When to Use — auto-triggers + explicit syntax

## Process — numbered steps (not prose)

## Output Format — template with placeholders

## Failure Handling — per-failure table

## Workflow Integration — which Phase (0-6)

## Handoff Protocol — next agent + action

Compliance Evaluation

After generating, check:

• mk: prefix in skill name
• Workflow phase anchoring (Phase 0-6 stated)
• Handoff protocol (next agent specified)
• Output format uses template with placeholders
• Failure handling covers identified failure modes
• SKILL.md ≤ 500 lines (per skill-authoring-rules.md Rule 3; overflow → references/ or step files)
• ## Gotchas section present (per Rule 1 — mandatory, placeholder acceptable day-1)
• No conflict with existing .claude/rules/

Score: X/8 — PASS (≥7) / FAIL (<7)

If FAIL: fix failing items before registering.

Output Format

## Skill Created: mk:{name}

**Directory:** .claude/skills/{name}/
**Files:** SKILL.md{, references/*.md}
**Compliance:** {X}/7 — {PASS|FAIL}
**Registered:** {yes|no — in SKILLS_ATTRIBUTION.md}

### Compliance Details
{checklist with ✓/✗ per item}

<!-- research-citation -->

Anthropic Skill-Building Lessons

1. Skills are folders — use scripts/, references/, assets/, lib/ creatively. See references/filesystem-patterns.md.
2. Gotchas = highest signal — every skill MUST have a ## Gotchas section. A day-1 placeholder is allowed; replace it as real failures emerge.
3. Progressive disclosure — SKILL.md stays under 500 lines per skill-authoring-rules.md Rule 3. Details live in references/ loaded on-demand. Step-file skills (with workflow.md) auto-pass.
4. Avoid railroading — describe outcomes, not step-by-step procedures. See references/good-vs-bad-examples.md.
5. Description = trigger condition — must answer "When should I use this?" not "What does this do?" Start with "Use when..."
6. On-demand hooks — skills can register session-scoped hooks in frontmatter for enforcement during execution.
7. config.json for setup — if skill needs user-specific values, store in config.json. Agent asks on first use.
8. Memory strategy — stateful skills should document what they persist and where.
9. Don't state the obvious — only include knowledge that pushes Claude beyond its defaults.
10. One skill type — classify using the 9-type taxonomy. See references/skill-types.md.

Mandatory checklist before finalizing any skill

• Description starts with "Use when..." (trigger condition)
• Has ## Gotchas section header (Rule 1 — placeholder acceptable day-1; grow over time)
• SKILL.md is under 500 lines (Rule 3 — or decomposed via references/ / step files)
• Steps are outcome-focused (no railroading)
• Uses filesystem beyond SKILL.md (or documents why not needed)

<!-- research-citation -->

• Classified into one Anthropic skill type
• Persistent state (if any) writes to ${CLAUDE_PLUGIN_DATA}, NOT skill directory (Rule 2)

References

• creation-workflow.md — Step 4: section-by-section guidance and examples
• skill-types.md — 9-type taxonomy with examples
• good-vs-bad-examples.md — writing descriptions, gotchas, and steps
• filesystem-patterns.md — when to use scripts/, references/, assets/, lib/, config.json

Failure Handling

Failure	Recovery
Name missing mk: prefix	Auto-prepend and warn
Duplicate skill name	Suggest alternative or update existing
Compliance check fails	List failing items, fix before registering

Gotchas

• Template must include ## Gotchas header — even as a placeholder. skill-authoring-rules.md Rule 1 is hard-enforced by validate-skill.py. Day-1 skills may use (none yet — grow from observed failures) but the header itself is mandatory.

<!-- research-citation -->

• Line cap is 500, not 150 — the canonical threshold comes from skill-authoring-rules.md Rule 3 (Anthropic progressive-disclosure guidance). Step-file skills (with workflow.md) auto-pass regardless of line count.
• Persistent state goes to ${CLAUDE_PLUGIN_DATA}, not skill directory — Rule 2 prevents data loss on plugin upgrade. framework-internal paths (.meowkit/memory/, session-state/) are exempt.