General Skill Maker

Create effective agent skills through guided briefing process that explores requirements, patterns, and best practices.

Workflow

6-step process: Brief → Load References → Plan Structure → Generate → Apply Best Practices → Present

Step 1: Briefing Questions

Ask user questions ONE AT A TIME (not all at once). Adapt based on answers.

Complete question library: See references/briefing-questions.md

Core Questions (REQUIRED):

1. Skill Purpose - "Jaki jest główny cel tego skilla?"
2. Usage Examples - "Podaj 2-3 przykłady jak użytkownik będzie prosił o użycie tego skilla?"
3. Environment/Platform - "Gdzie będzie używany ten skill?"
4. Freedom Level - "Jak deterministyczny ma być proces?"
5. Expected Output - "Jaki jest oczekiwany output?"

Optional Questions (ask based on context): 6. Task Scope - Single vs related tasks 7. User Base - Solo vs team 8. Scripts Needed - For repetitive operations 9. Reference Documentation - API docs, schemas, examples 10. Assets/Templates - Templates, boilerplate code 11. External Integrations - APIs, databases, tools 12. Workflow Type - Multi-step vs single-step 13. User Interaction - Interactive vs autonomous 14. Quality Standards - Validation requirements

Briefing strategy:

• Start with core questions (1-5)
• Ask optional questions (6-14) only when relevant based on previous answers
• Adapt and skip irrelevant questions
• See references/briefing-questions.md for detailed questions, follow-ups, and examples

Step 2: Load Appropriate References

Based on user's answers, read relevant reference files:

Always read:

• references/best-practices.md

  - Universal principles for all skills

Read based on workflow type:

• references/workflow-patterns.md

  - If multi-step workflow with sequential logic

• references/integration-patterns.md

  - If external API/tool integrations

• references/tool-based-patterns.md

  - If specific tools/scripts dominate

• references/domain-knowledge-patterns.md

  - If primarily about conveying expertise

Load only what's needed - Don't read all references unnecessarily.

Step 3: Plan Structure

Based on briefing answers, determine skill structure:

Required:

• SKILL.md

  - Always required

Optional folders to create:

• scripts/

  - If Q6 answered yes (executable code)

• references/

  - If Q7 answered yes (detailed documentation)

• assets/

  - If Q8 answered yes (templates, files for output)

Plan SKILL.md sections:

Frontmatter (always):

---
name: skill-name
description: Comprehensive description with WHEN to use
---

Body sections (adapt based on needs):

• Overview/Introduction
• Quick Start (if complex)
• Step-by-step instructions (if multi-step)
• Examples (always recommended)
• Edge cases (if complex)
• References to supporting files (if using scripts/references/assets)
• Quality checks (if validation needed)

Important considerations:

• Keep SKILL.md under 500 lines
• Move detailed content to references/
• Structure over prose
• Include concrete examples
• Clear section headings

Step 4: Generate Skill

Create the actual skill files based on plan.

A. Create Directory Structure

mkdir -p skill-name/{scripts,references,assets}

Only create folders that are needed (from Step 3 plan).

B. Write SKILL.md

Frontmatter Guidelines:

---
name: skill-name  # hyphen-case, max 64 chars, no leading/trailing hyphens
description: >
  [What it does]. [When to use it]. Use when [triggers].
  Trigger on phrases like "[example1]", "[example2]", "[example3]".
---

Description best practices:

• Include WHAT it does (capabilities)
• Include WHEN to use (contexts, triggers)
• Include specific KEYWORDS users naturally say
• Use 1-1024 characters wisely
• Think semantic matching - agent uses language understanding

Body structure:

Use imperative/infinitive form (e.g., "Run tests", not "Running tests" or "The agent should run tests").

# Skill Name

[One-line description]

## Quick Start (optional for complex skills)

[Minimal example to get started fast]

## Core Workflow

[Step-by-step instructions]

1. [Step 1]
2. [Step 2]
3. [Step 3]

## Examples

[Concrete examples with input → output]

## Supporting Resources (if applicable)

- **scripts/script.py** - [What it does]
- **references/details.md** - [What it contains, when to load]
- **assets/template.html** - [What it is, how to use]

## Quality Checks (if applicable)

[Validation steps, success criteria]

Writing guidelines:

• ✅ Structured (headers, lists, steps)
• ✅ Concrete examples
• ✅ Only add what LLM doesn't already know
• ✅ Imperative form
• ❌ Prose/walls of text
• ❌ Generic explanations LLM knows
• ❌ Unnecessary details

C. Create Scripts (if needed)

For each script identified in briefing:

#!/usr/bin/env python3
"""
Script description.

Usage:
    script.py [args]
"""

# Implement with:
# - Clear error messages
# - Input validation
# - Graceful error handling
# - Self-contained or document dependencies

Test scripts:

• Run scripts to verify they work
• Test with edge cases
• Ensure error messages are helpful

D. Create References (if needed)

For each reference file identified:

# Reference Title

[Focused documentation loaded on-demand]

## Section 1

[Content]

## Section 2

[Content]

Reference guidelines:

• Keep focused (one topic per file)
• Use table of contents for 100+ lines
• Link from SKILL.md with clear "when to read" guidance
• Split domain-specific content (e.g., finance.md, sales.md)

E. Create Assets (if needed)

Copy or create template files:

• Document templates
• Boilerplate code
• Images/diagrams
• Config examples

Asset guidelines:

• Not loaded into context
• Used in agent's output
• Document in SKILL.md how to use them

Step 5: Apply Best Practices

Review generated skill against quality criteria.

Quick checklist:

1. Description - Keywords, triggers, when to use? ✓
2. Length - SKILL.md under 500 lines? ✓
3. Structure - Headers, steps, lists, examples? ✓
4. Content - Specific, actionable, no LLM basics? ✓
5. Progressive Disclosure - Details in references/? ✓
6. Supporting Files - Scripts tested, references organized? ✓

Comprehensive quality validation:

See references/quality-checklist.md for:

• Complete 25-point quality checklist
• Critical/High Quality/Excellence criteria
• Scoring system and validation commands
• Common quality issues and fixes

If any check fails: Revise before presenting to user.

Step 6: Present and Iterate

1. Show generated skill structure to user
2. Display SKILL.md content
3. Explain key sections and design decisions
4. Show any scripts/references/assets created
5. Ask: "Czy chcesz jakieś zmiany lub doprecyzowania?"
6. Iterate based on feedback
7. Save all files to skill directory

Validation

Before finalizing, verify:

• Name is hyphen-case, max 64 chars, no leading/trailing hyphens
• Description is 1-1024 chars, includes what/when/triggers
• SKILL.md under 500 lines
• Structured format (headers, lists, examples)
• Scripts tested (if any)
• References organized (if any)
• No hardcoded absolute paths
• User confirms it matches their needs

Tips for Success

1. Start simple, iterate - Begin with minimal viable skill, add based on real usage
2. User-specific language - If user speaks Polish, use Polish in generated skill content
3. Progressive disclosure - Keep SKILL.md lean, details in references/
4. Concrete over abstract - Examples > explanations
5. Test and refine - Encourage user to test and report issues

See complete workflow example: references/complete-example.md shows end-to-end process from briefing to final PDF processing skill.

Common Mistakes to Avoid

Quick list (see references/common-mistakes.md for detailed guide):

• Too generic description - "Helps with PDFs" → "Extracts text from PDFs, fills forms, merges documents"
• Too long SKILL.md - 2000+ lines → Split into references/
• Missing examples - Abstract descriptions → Concrete input/output examples
• Hardcoded paths -

  /Users/john/...

  → Relative paths or ${baseDir}
• Wall of text - Prose → Structured format (headers, lists, steps)
• Unnecessary scripts - Don't create scripts agent can easily write
• Everything in one file - Monolithic SKILL.md → Progressive disclosure

See references/common-mistakes.md for:

• Complete anti-patterns guide (Encyclopedia, Everything Bagel, Secret Handshake, etc.)
• Before/after refactoring examples
• Prevention checklist

Advanced Features (Optional)

For advanced features like hooks, subagent context, allowed-tools, and model override:

See references/advanced-features.md

Features include:

• Hooks - Lifecycle automation (PreToolUse, PostToolUse validation)
• Subagent Context - Isolated execution with specific tools
• Allowed Tools - Pre-approve tools to skip permission prompts
• Dynamic Context - Inject fresh external data before execution
• Model Override - Use haiku/opus for specific skills

Note: Most skills don't need advanced features. Use only when they solve specific problems.