skill-creator

📋 Skill Creator = Knowledge Framework Application

Core Principle: Skills are modular packages extending Claude's capabilities through specialized workflows, domain knowledge, and bundled resources. Every skill follows Knowledge Framework principles - MECE structure, Mermaid diagrams, minimal verbosity, and Ground Truth attribution.

Skill Package Anatomy (Continuant - TD):

graph TD
    Skill[Skill Package] --> YAML[YAML Frontmatter]
    Skill --> Body[Markdown Body]
    Skill --> Resources[Bundled Resources]

    YAML --> Name[name: required]
    YAML --> Desc[description: required]
    YAML --> Tools[allowed-tools: optional]

    Body --> Principle[Core Principle]
    Body --> Diagrams[Mermaid TD/LR]
    Body --> Sections[MECE Sections]

    Resources --> Scripts[scripts/]
    Resources --> Refs[references/]
    Resources --> Assets[assets/]

Skill Creation & Evolution Workflow (Occurrent - LR):

graph LR
    A[Understand Use Cases] --> B[Plan Resources]
    B --> C[Init Skill Template]
    C --> D[Edit with KF Principles]
    D --> E[Validate]
    E --> F[Package if Complete]
    F --> G[Iterate]
    G --> D

    H[Runtime Error] --> I[Update SKILL.md/Code]
    I --> J[Increment Version]
    J --> K[Log in SKILL_RELEASE_LOG.md]
    K --> E

Ontological Rule: TD for skill structure (what exists), LR for creation/evolution process (what happens)

Primary source: skill-creator/SKILL.md (Anthropic base + Knowledge Framework integration) Original session: e9ce3592-bd66-4a98-b0e7-fcdd8edb5d42 by Daniel Kravtsov (2025-11-13) - v2.0.0 KF migration Latest update: e9ce3592-bd66-4a98-b0e7-fcdd8edb5d42 by Daniel Kravtsov (2025-11-13) - v2.2.0 Mermaid headers Release log: See SKILL_RELEASE_LOG.md for full version history

🎯 What Skills Provide

¶1 Skills transform Claude into specialized agent with domain expertise no model can fully possess.

¶2 Four core capabilities:

• Specialized workflows - Multi-step procedures for specific domains
• Tool integrations - Instructions for file formats, APIs, databases
• Domain expertise - Company knowledge, schemas, business logic
• Bundled resources - Scripts, references, assets for complex tasks

¶3 Progressive disclosure - Three-level loading:

1. Metadata (name + description) - Always in context (~100 words)
2. SKILL.md body - When skill triggers (<5k words)
3. Bundled resources - As needed (unlimited, scripts executed without reading)

📐 Skill Structure (Anatomy)

¶1 Required: SKILL.md with YAML frontmatter + Markdown body

YAML Frontmatter (Required Fields):

---
name: skill-identifier  # lowercase, hyphens only, max 64 chars, no "anthropic"/"claude"
description: What skill does + when to use + trigger phrases. Max 1024 chars. No angle brackets.
---

YAML Optional Fields:

• allowed-tools - Restricts tools (e.g., Read, Write, Glob, Grep)
• model - Specific model or "inherit" (default)
• version - Version tracking (e.g., "2.0.0")
• disable-model-invocation: true - Manual-only invocation via /skill-name
• mode: true - Appears in "Mode Commands" section

CRITICAL: description field is PRIMARY field Claude uses for skill activation. Include trigger phrases directly: "Use when user says 'create PDF', 'extract text', 'merge documents', or works with PDF files."

¶2 Optional: Bundled Resources

skill-name/
├── SKILL.md (required)
└── Bundled Resources (optional)
    ├── scripts/          - Executable code (Python/Bash/etc.)
    ├── references/       - Documentation loaded as needed
    └── assets/           - Files used in output (templates, icons)

Scripts (scripts/) - When same code rewritten repeatedly or deterministic reliability needed

• Example: scripts/rotate_pdf.py
• Token efficient, executed without loading into context

References (references/) - Documentation Claude references while working

• Examples: references/schema.md, references/api_docs.md
• Loaded only when Claude determines it's needed
• Avoid duplication: Info lives in SKILL.md OR references, not both

Assets (assets/) - Files used in final output, not loaded into context

• Examples: assets/logo.png, assets/template.html
• Templates, boilerplate, images copied or modified

🔄 Skill Creation Process

¶1 Step 1: Understand Use Cases

Ask user for concrete examples:

• "What functionality should this skill support?"
• "Can you give examples of how this skill would be used?"
• "What would a user say that should trigger this skill?"

Skip only if usage patterns already clearly understood.

¶2 Step 2: Plan Resources

For each example, identify reusable resources:

• Scripts - Code rewritten repeatedly? → scripts/process.py
• References - Schemas/docs needed repeatedly? → references/schema.md
• Assets - Templates/boilerplate? → assets/template.html

¶3 Step 3: Init Skill Template

scripts/init_skill.py <skill-name> --path <output-directory>

Creates skill directory with SKILL.md template and example resource directories. Delete unused examples.

Skip if skill already exists (go to Step 4).

¶4 Step 4: Edit Skill with Knowledge Framework

CRITICAL: Apply Knowledge Framework to SKILL.md

Every SKILL.md MUST follow Knowledge Framework principles. See knowledge-framework skill for complete documentation framework details.

Required elements:

1. Core Principle - One sentence defining skill's essence
2. Mermaid Diagrams with Headers - Each diagram needs descriptive header (2-5 words specific to THIS skill)
   • TD diagram for structure (Continuant)
   • LR diagram for process (Occurrent)
3. Ontological Rule - TD for Continuants, LR for Occurrents
4. Ground Truth Attribution - Session ID + primary source
5. MECE Sections - Emoji-sections (🎯, 📐, 🔄) with ¶-numbering
6. Minimal Verbosity - No unnecessary words, DRY principle

Template Structure:

---
name: skill-name
description: What + when + triggers. Max 1024 chars.
---
## 📋 [Skill Name] = [Core Domain]

**Core Principle:** [One sentence - what skill does and principles it follows]

**[Specific Description] (Continuant - TD):**
```mermaid
graph TD
    [Define structure nodes]

[Specific Description] (Occurrent - LR):

graph LR
    [Define process flow]

Ontological Rule: TD for [structure], LR for [process]

Primary source: [Original documentation/base] Session ID: [get_session_id.py output] by [Author] (YYYY-MM-DD)

🎯 [Section 1 Name]

¶1 [Content with ¶-numbering]

📐 [Section 2 Name]

¶1 [Content]

🔄 [Section 3 Name]

¶1 [Content]

**Example headers:**
- ✅ `**Skill Package Anatomy (Continuant - TD):**` (specific to skill-creator)
- ✅ `**Email Processing Workflow (Occurrent - LR):**` (specific to email-parser)
- ❌ `**System Structure:**` (too generic)


**Writing Style:**
- **Imperative/infinitive form** (verb-first): "To accomplish X, do Y"
- NOT second person ("you should")
- Objective, instructional language

**Portability (CRITICAL):**

✅ **DO use relative paths:**
- `scripts/process.py` - Within skill directory
- `references/schema.md` - Within skill directory
- `algorithms/data/config.json` - Relative to repository root

❌ **DO NOT use absolute paths:**
- `/Users/username/projects/file.md` ❌
- `~/projects/repo/data/` ❌
- `C:\Users\username\file.txt` ❌

**YAML Examples:**

✅ **With triggers in description:**
```yaml
---
name: pdf-processor
description: Extract text, fill forms, merge PDFs. Use when user mentions "PDF", "extract from document", "fill form", or works with PDF files.
---

✅ Bilingual triggers:

---
name: data-analyzer
description: Analyze customer data. Use when user says "analyze data", "generate report", "анализируй данные", "создай отчет", or works with analytics.
allowed-tools: Read, Bash, Glob, Grep
---

✅ Manual-only invocation:

---
name: database-migration
description: Migrate production database. DANGER: Manual invocation only via /database-migration.
disable-model-invocation: true
allowed-tools: Bash, Read, Write
---

¶5 Step 5: Validate & Package

Development validation (fast):

scripts/quick_validate.py <path/to/skill-folder>

Package only when COMPLETE and ready to distribute:

scripts/package_skill.py <path/to/skill-folder>

Packaging automatically validates first, then creates distributable zip.

❌ DO NOT package during:

• Incremental development
• Testing/iterating
• User still editing

✅ DO package when:

• Skill COMPLETE and ready to distribute
• Final version before sharing externally
• User explicitly requests packaged skill

¶6 Step 6: Iterate & Self-Healing

After testing, identify improvements:

1. Use skill on real tasks
2. Notice struggles or inefficiencies
3. Update SKILL.md or bundled resources
4. Test again

CRITICAL: Self-Healing Skills (Runtime Error Recovery)

When skill encounters error during execution, skill should automatically update itself:

1. Identify error - Script fails, instruction unclear, resource missing
2. Fix the issue - Update SKILL.md, fix script, add missing reference
3. Increment version - Update version in YAML frontmatter (e.g., "2.0.0" → "2.0.1")
4. Log change - Add entry to SKILL_RELEASE_LOG.md
5. Preserve attribution - ALWAYS keep original session ID + author, add new session for update

SKILL_RELEASE_LOG.md Format:

# Skill Release Log

## [2.0.1] - 2025-11-13
**Session ID:** `abc123def456` by Daniel Kravtsov
**Type:** Bugfix
**Changes:**
- Fixed script path in references section
- Added missing error handling in `scripts/process.py`
**Issue:** Runtime error when processing empty file

## [2.0.0] - 2025-11-10
**Session ID:** `xyz789abc123` by Daniel Kravtsov
**Type:** Major update
**Changes:**
- Migrated to Knowledge Framework structure
- Added Mermaid diagrams
- Restructured with MECE sections
**Issue:** Initial release with KF compliance

Self-Healing Rules:

• ✅ DO preserve original Ground Truth attribution (session ID + author)
• ✅ DO add new session entry in SKILL_RELEASE_LOG.md
• ✅ DO increment version (patch: x.x.1, minor: x.1.0, major: 1.0.0)
• ❌ DO NOT remove existing session IDs from Ground Truth
• ❌ DO NOT overwrite SKILL_RELEASE_LOG.md - only append

Example - Preserving attribution:

**Primary source:** `skill-creator/SKILL.md` (Anthropic base + Knowledge Framework integration)
**Original session:** `e9ce3592-bd66-4a98-b0e7-fcdd8edb5d42` by Daniel Kravtsov (2025-11-13)
**Latest update:** `abc123def456` by Daniel Kravtsov (2025-11-15) - Fixed runtime error

🔗 Knowledge Framework Integration

¶1 Every skill demonstrates Knowledge Framework principles:

When creating ANY skill, the skill itself should follow these patterns (see knowledge-framework skill for full details):

Required Elements:

• Thesis/Core Principle - Single sentence defining skill essence
• Mermaid Diagrams - Minimum 1, complex skills need both TD and LR
• Ontological Rule - TD for structure (nouns), LR for process (verbs)
• MECE Sections - 3-7 emoji-sections with ¶-numbering
• Ground Truth - Session ID + primary source + dates

DRY Principle:

• Each fact appears exactly ONCE
• Use references: "See references/schema.md for table structure"
• No duplication between SKILL.md and references files

Minimal Verbosity:

• Use minimum tokens sufficient for execution/understanding
• Remove unnecessary examples, verbose explanations
• Keep only essential procedural knowledge in SKILL.md

¶2 Reference the knowledge-framework skill:

When editing SKILL.md, if structure is complex or unclear, reference:

**See:** `knowledge-framework` skill for MECE/BFO principles

¶3 Common anti-patterns to avoid:

❌ No Core Principle or thesis statement ❌ Missing Mermaid diagrams ❌ Long verbose descriptions without structure ❌ Absolute paths (/Users/, /home/, C:\) ❌ Duplication between SKILL.md and references ❌ Missing Ground Truth attribution (session ID, source) ❌ Flat structure without emoji-sections and ¶-numbering

✅ Skill Quality Checklist

Before finalizing skill:

Structure:

• Core Principle (1 sentence defining skill)
• Mermaid diagrams (TD for structure, LR for process)
• Diagram headers (2-5 words, specific to THIS skill, includes type/direction)
• Ontological Rule statement
• MECE emoji-sections (3-7 sections)
• ¶-numbering within sections
• Ground Truth attribution (session ID + source)

YAML Frontmatter:

• name field (lowercase, hyphens, max 64 chars)
• description includes trigger phrases (max 1024 chars)
• No angle brackets in description
• Optional fields correct (allowed-tools, version, etc.)

Portability:

• No absolute paths (/Users/, ~/, C:\)
• Relative paths for all file references
• Works across different machines without modification

Resources:

• Scripts in scripts/ (if needed)
• References in references/ (if needed)
• Assets in assets/ (if needed)
• No duplication between SKILL.md and references
• Example files removed if not needed

Quality:

• Imperative/infinitive writing style
• Minimal verbosity (no unnecessary words)
• DRY principle (no duplication)
• Validated with quick_validate.py

Versioning & Evolution:

• version field in YAML frontmatter
• SKILL_RELEASE_LOG.md exists if skill has updates
• Original session ID preserved in Ground Truth
• Latest update session documented (if applicable)
• Version incremented correctly (patch/minor/major)

---

Meta Note: This skill itself demonstrates Knowledge Framework - Core Principle, Mermaid diagrams (TD structure + LR process), emoji-sections with ¶-numbering, Ground Truth attribution, minimal verbosity. Skills are self-healing - they update themselves on runtime errors with proper versioning and release log tracking.