🧠 Skill Generator Ultra

The most comprehensive AI Skill creation toolkit on the market.
Bộ công cụ tạo AI Skill toàn diện nhất trên thị trường.

---

🌍 Language / Ngôn ngữ

• English
• Tiếng Việt

---

English

📋 Table of Contents

• Introduction
• Key Features
• Installation
• Usage
• Scripts & Tools
• Slash Commands
• Project Structure
• Platform Compatibility
• Evaluation System
• Documentation
• FAQ
• Changelog
• Contributing
• License

---

Introduction

Skill Creator Ultra is a specialized AI Skill for Google Antigravity (and 7 other platforms). It helps you create production-quality AI Skills through an 8-phase pipeline — even if you DON'T know what a skill is, what YAML is, or how to write SKILL.md.

You only need:

• ✅ An idea about what you want to automate
• ✅ A workflow / process / logic in your head

AI handles the rest:

• 🎤 Smart interview with 5 extraction techniques + Quick Mode
• 🔍 Auto-detect complexity → choose the right architecture
• 🏗️ Generate a complete skill package for 7 platforms
• 🧪 Test, evaluate with 7-dimension scoring, scan for security issues
• 📦 Package and publish to marketplaces

What makes Ultra different from skill-generator?

Ultra merges the best of two skill creators and adds competitive features:

Feature	Source	Description
🎤 8-Phase Pipeline	skill-generator + Anthropic	Interview → Extract → Detect → Generate → Test → Eval → Iterate → Optimize
⚡ Quick Mode	NEW (GPTs Builder-inspired)	One-shot skill creation for clear descriptions
📊 7-Dimension Eval	NEW (DeepEval-inspired)	Correctness, Completeness, Format, Adherence, Safety, Efficiency, Robustness
🔐 Security Scanning	NEW (Promptfoo-inspired)	5 checks: Prompt Injection, PII, Secrets, Scope, Destructive Commands
🔄 CI/CD Integration	NEW	ci_eval.py script for GitHub Actions / GitLab CI
📦 Package & Publish	NEW	.skill packaging → publish to GitHub, Agent Skills Market, SkillsMP
🤖 3 AI Agents	Anthropic	Grader + Comparator + Analyzer for blind A/B testing
🎯 Description Optimization	Anthropic	Trigger Eval → Pushy descriptions → anti-triggers
🌐 Multi-Platform Export	skill-generator	Antigravity, Claude, Cursor, Windsurf, OpenClaw, Cline, Copilot
🔎 Pattern Detection	skill-generator	Auto-detect complexity → choose architecture
🔗 System Mode	skill-generator	Multi-skill pipeline orchestration with I/O contracts
📝 16 Prompt Principles	Both	16 principles with ❌/✅ code examples
❌ 15 Anti-Patterns	skill-generator	Named patterns: 🐙 Octopus, 👻 Ghost, 🎪 Circus...
✍️ 10 Writing Patterns	NEW	Expert patterns: 🎯 Sniper, 🧬 DNA, 🪞 Mirror...
📺 Eval Viewer	Anthropic	HTML viewer for benchmark results

Who is this for?

Audience	Example
Non-coders	Sales rep who wants AI to auto-draft price quotes
Developers	Want AI to format commit messages, review code, generate docs
Team Leads	Standardize team workflows as reusable, testable AI Skills
Skill Developers	Build & publish skills to marketplaces for others to use
Prompt Engineers	Upgrade from ad-hoc prompts to structured, evaluated skill systems

---

✨ Key Features

Pipeline — 8 Phases

Phase 1-5: CREATE (always runs)
───────────────────────────────────────
  1. 🎤 Interview  — Smart extraction + Quick Mode
  2. 🔬 Extract    — Raw info → structured components
  3. 🔎 Detect     — Pattern detection + complexity scoring
  4. 🏗️ Generate   — Full package for 7 platforms
  5. 🧪 Test       — Dry run + validation + Package & Publish

Phase 6-8: REFINE (optional, for production skills)
───────────────────────────────────────
  6. 📊 Eval       — 7-dimension scoring + security scanning
  7. 🔄 Iterate    — Fix → re-test → blind compare
  8. 🎯 Optimize   — Trigger accuracy tuning

5 Operating Modes

Mode	When	Phases
⚡ Quick Mode	User describes clearly (triggers + steps + rules + output)	4 → 5
Standard	User has rough idea, needs some clarification	1 (short) → 3 → 4 → 5
Full Interview	User only knows "I want automation"	1 → 2 → 3 → 4 → 5
System Mode	Multi-step workflow (≥3 independent steps)	1 → 2 → 3 → 4S → 5
Improve Mode	Existing skill needs eval + improvement	6 → 7

---

🔧 Installation

⚡ One-Click Install (Recommended)

Install with a single command — the script auto-detects your OS, lets you choose a platform, clones the repo, and creates all bridge/rule files:

macOS / Linux:

curl -sL https://raw.githubusercontent.com/marketingjuliancongdanh79-pixel/skill-generator/main/install.sh | bash

Windows (PowerShell):

irm https://raw.githubusercontent.com/marketingjuliancongdanh79-pixel/skill-generator/main/install.ps1 | iex

💡 The installer shows an interactive menu to choose your platform (Antigravity, Claude Code, Cursor, Windsurf, Cline, Copilot, OpenClaw) or install ALL at once.

---

Prerequisites (Manual Install)

• A supported AI Agent platform (see compatibility table below)
• Python 3.8+ (optional, for scripts)

🟢 Google Antigravity (Native — Recommended)

Global Installation — Available in ALL projects:

# macOS / Linux
cp -r skill-creator-ultra ~/.gemini/antigravity/skills/skill-creator-ultra

# Windows (PowerShell)
Copy-Item -Recurse skill-creator-ultra "$env:USERPROFILE\.gemini\antigravity\skills\skill-creator-ultra"

Workspace Installation — Only in current project:

cp -r skill-creator-ultra .agent/skills/skill-creator-ultra

After install: Open a new chat and say "tạo skill mới" to verify.

---

🟣 Claude Code (Anthropic)

Claude Code supports Custom Commands via .claude/commands/.

Step 1: Copy the skill

# Global
cp -r skill-creator-ultra ~/.claude/commands/skill-creator-ultra

# Or Workspace
mkdir -p .claude/commands
cp -r skill-creator-ultra .claude/commands/skill-creator-ultra

Step 2: Create bridge file — append to CLAUDE.md at project root:

## Skill Creator Ultra
When user asks to "create skill", "turn workflow into skill", or "automate this":
- Read `.claude/commands/skill-creator-ultra/SKILL.md`
- Follow the 8 Phase pipeline
- Reference resources/ for templates and best practices

Step 3: Use it

You: Create a skill for formatting commit messages
Claude: (Reads SKILL.md → Starts interviewing)

---

🔵 Cursor

Cursor supports Rules — use skill-creator-ultra as a Custom Rule.

Step 1: Copy into Cursor Rules

mkdir -p .cursor/rules
cp -r skill-creator-ultra .cursor/rules/skill-creator-ultra

# Windows
New-Item -ItemType Directory -Force -Path ".cursor\rules"
Copy-Item -Recurse skill-creator-ultra ".cursor\rules\skill-creator-ultra"

Step 2: Create rule file .cursor/rules/skill-creator-ultra.mdc:

---
description: Create AI Skills from ideas or workflows using 8-phase pipeline
globs:
alwaysApply: false
---

When user requests "create skill", "turn workflow into skill", or "automate":
- Read `.cursor/rules/skill-creator-ultra/SKILL.md`
- Follow 8 Phase pipeline
- Reference resources/ for templates and best practices

Step 3: Use it

You: @skill-creator-ultra create a skill for weekly reports
Cursor: (Reads rule → Applies SKILL.md → Starts interview)

---

🟠 Windsurf / Codeium

Windsurf uses Cascade Rules — similar to Cursor.

mkdir -p .windsurf/rules
cp -r skill-creator-ultra .windsurf/rules/skill-creator-ultra

Create .windsurf/rules/skill-creator-ultra.md:

---
trigger: manual
description: Create AI Skills from ideas or workflows
---

When user requests "create skill" or "automate this":
- Read `.windsurf/rules/skill-creator-ultra/SKILL.md`
- Follow 8 Phase pipeline

---

🟤 Cline (VS Code Extension)

Cline uses Custom Instructions — paste SKILL.md into System Prompt.

cp -r skill-creator-ultra .clinerules/skill-creator-ultra

In Cline Settings → Custom Instructions, add:

When user requests "create skill" or "automate":
- Read .clinerules/skill-creator-ultra/SKILL.md
- Follow 8 Phase pipeline

---

⚫ GitHub Copilot

cp -r skill-creator-ultra .github/skill-creator-ultra

Add to .github/copilot-instructions.md:

## Skill Creator Ultra
When user requests "create skill" or "automate this":
- Read `.github/skill-creator-ultra/SKILL.md`
- Follow 8 Phase pipeline

⚠️ Note: Copilot doesn't support running scripts directly.

---

🐾 OpenClaw (AI Gateway)

OpenClaw uses System Prompt — paste SKILL.md content.

1. Open agent config in OpenClaw Admin Panel
2. Copy contents of SKILL.md into System Prompt
3. Optionally upload resources/ files to knowledge base
4. Test: "Tạo cho tao 1 skill tự soạn email báo giá"

⚠️ OpenClaw has token limits. If too long, use slim SKILL.md only.

---

📥 Quick Install Table

Platform	Command (macOS/Linux)
Antigravity (Global)	cp -r skill-creator-ultra ~/.gemini/antigravity/skills/
Antigravity (Workspace)	cp -r skill-creator-ultra .agent/skills/
Claude Code	cp -r skill-creator-ultra ~/.claude/commands/
Cursor	cp -r skill-creator-ultra .cursor/rules/
Windsurf	cp -r skill-creator-ultra .windsurf/rules/
Cline	cp -r skill-creator-ultra .clinerules/
Copilot	cp -r skill-creator-ultra .github/
OpenClaw	Copy SKILL.md content → Agent System Prompt

---

🚀 Usage

Way 1: Natural language (AI auto-triggers)

What you say	AI understands
"Create a new skill for deploy automation"	→ Activates skill-creator-ultra
"Turn this review workflow into a skill"	→ Activates skill-creator-ultra
"I want AI to auto-write weekly reports"	→ Activates skill-creator-ultra
"Make a new skill" / "Tạo skill"	→ Activates skill-creator-ultra
"Improve my existing skill X"	→ Activates skill-creator-ultra (Improve Mode)

Way 2: Slash commands

Type / in your AI chat to access quick commands (see Slash Commands below).

Full Workflow

You describe your idea
    ↓
⚡ Quick Mode check — is description clear enough?
    ├── YES → Confirm → Generate skill immediately
    └── NO → Smart interview (5-10 questions)
         ↓
    🔍 AI detects pattern + calculates complexity score
         ↓
    🏗️ AI generates full skill package (SKILL.md + resources + examples + scripts)
         ↓
    🧪 AI dry-runs with realistic test scenarios
         ↓
    ✅ Skill complete, ready to deploy!
    
    (Optional: Phase 6-8 for production refinement)
         ↓
    📊 7-dimension eval + security scan
         ↓
    🔄 Iterate based on scores
         ↓
    🎯 Optimize trigger accuracy

Real-world Example

You: "I want a skill that auto-writes weekly reports. When I say my tasks, it formats them into 4 sections: Done, In Progress, Blockers, Next Week. Keep under 400 words because my boss reads on mobile."

AI: This is a Quick Mode trigger (you gave: trigger + steps + rules + output format). Let me confirm and generate...

→ Generates weekly-report-writer skill with SKILL.md, 2 examples (happy path + edge case), constraints, deployed in 2 minutes.

---

🔧 Scripts & Tools (9 Utilities)

# CI/CD evaluation checker (NEW)
python scripts/ci_eval.py eval_results.json --min-score 85
python scripts/ci_eval.py eval_results.json --strict --format json

# Package skill for distribution (NEW)
python scripts/package_skill.py ./my-skill/ ./dist/my-skill.skill

# Validate SKILL.md structure
python scripts/validate_skill.py ./path/to/my-skill/

# Simulate dry-run
python scripts/simulate_skill.py ./path/to/my-skill/

# Audit against 7 Perfect Principles → S/A/B/C/D/F grade
python scripts/skill_audit.py ./path/to/my-skill/
python scripts/skill_audit.py ./path/to/my-skill/ --json --strict

# View stats + Cognitive Load Score
python scripts/skill_stats.py ./path/to/my-skill/

# Export to Cursor/Claude/Windsurf/Cline/Copilot/OpenClaw
python scripts/skill_export.py ./path/to/my-skill/ --platform cursor
python scripts/skill_export.py ./path/to/my-skill/ --platform all

# Compare 2 versions of a skill
python scripts/skill_compare.py ./old-skill/ ./new-skill/

# Scaffold a new skill from scratch
python scripts/skill_scaffold.py my-new-skill --full
python scripts/skill_scaffold.py my-new-skill --interactive

Script Details

Script	Purpose	Exit Codes
ci_eval.py	CI/CD pipeline checker — validates eval score + security	0=pass, 1=iterate, 3=security-block
package_skill.py	Create .skill distribution archive	0=success
validate_skill.py	Check SKILL.md YAML, sections, structure	0=valid, 1=errors
skill_audit.py	Audit 7 principles → grade S/A/B/C/D/F	0=pass
skill_scaffold.py	Generate skeleton with correct structure	0=created
skill_export.py	Export to 7 platforms	0=exported
skill_compare.py	Diff 2 versions, highlight changes	0=compared
skill_stats.py	Statistics + Cognitive Load score	0=analyzed
simulate_skill.py	Dry-run simulation	0=simulated

---

⚡ Slash Commands (8 Commands)

Type / in your AI chat to access:

Command	Description
/skill-generate	🧠 Create new skill from idea (8-Phase pipeline)
/skill-audit	🔍 Audit skill against 7 principles
/skill-export	📦 Export to other platforms
/skill-stats	📊 View stats + Cognitive Load
/skill-compare	🔄 Compare 2 skill versions
/skill-scaffold	🧩 Create new skill skeleton
/skill-validate	✅ Check SKILL.md validity
/skill-simulate	🧪 Simulate dry-run

---

📁 Project Structure

skill-creator-ultra/                    56 files | 193 KB resources
├── SKILL.md                            🧠 Main orchestrator (388 lines, 8 phases)
├── README.md                           📖 This file
│
├── phases/                             🎬 8 phase guides
│   ├── phase1_interview.md             🎤 Smart interview + ⚡ Quick Mode
│   ├── phase2_extract.md               🔬 Knowledge extraction → components
│   ├── phase3_detect.md                🔎 Pattern detection + complexity scoring
│   ├── phase4_generate.md              🏗️ 7-platform code generation (545 lines)
│   ├── phase5_test.md                  🧪 Dry run + validation + 📦 Package & Publish
│   ├── phase6_eval.md                  📊 7-dimension eval + 🔐 security scanning
│   ├── phase7_iterate.md               🔄 Fix → re-test → blind compare
│   └── phase8_optimize.md              🎯 Description trigger optimization
│
├── agents/                             🤖 Eval agents (from Anthropic)
│   ├── grader.md                       Assertion grading (5-point scale)
│   ├── comparator.md                   Blind A/B comparison
│   └── analyzer.md                     Root cause analysis
│
├── resources/                          📚 17 reference docs (193 KB)
│   ├── skill_writing_guide.md          ✍️ 10 expert patterns (634 lines) — 🎯🧬🪞🏗️🛡️🎚️🔬🌳📐📆
│   ├── prompt_engineering.md           🧠 16 principles (532 lines) — specificity, CoT, few-shot...
│   ├── anti_patterns.md                ❌ 15 named patterns (377 lines) — 🐙🎭💨🔓📜🤖🎪...
│   ├── eval_guide.md                   📊 7-dimension eval reference (145 lines)
│   ├── description_optimization.md     🎯 Trigger accuracy + anti-triggers (91 lines)
│   ├── skill_template.md              📄 Standard SKILL.md template + checklist
│   ├── interview_questions.md          🎤 30+ questions + Expert Probing
│   ├── industry_questions.md           🏭 8 industry-specific question sets
│   ├── pattern_detection.md            🔍 Decision tree + Complexity scoring
│   ├── advanced_patterns.md            🧩 6 architecture patterns
│   ├── composition_cookbook.md          🔗 5 multi-skill composition patterns
│   ├── checklist.md                    ✅ 2-tier quality checklist (Basic + Expert)
│   ├── schemas.md                      📋 JSON schemas for eval
│   ├── scripts_guide.md               🔧 Scripts usage guide
│   ├── blueprints.md                   📘 10 ready-made skill templates
│   ├── versioning_guide.md             📆 Skill upgrade guide
│   └── script_integration.md           ⚙️ Script integration (7 security layers)
│
├── scripts/                            🔧 9 Python scripts
│   ├── ci_eval.py                      🆕 CI/CD evaluation checker
│   ├── package_skill.py                🆕 .skill file packager
│   ├── validate_skill.py               Validate SKILL.md structure
│   ├── simulate_skill.py               Simulate dry-run
│   ├── skill_audit.py                  Audit 7 principles → grade S-F
│   ├── skill_export.py                 Export to 7 platforms
│   ├── skill_stats.py                  Statistics + Cognitive Load
│   ├── skill_compare.py                Compare 2 versions
│   └── skill_scaffold.py              Scaffold new skill
│
├── examples/                           🎯 5 snapshot examples
│   ├── example_git_commit.md           🟢 Simple skill (Complexity ≤ 5)
│   ├── example_api_docs.md             🟡 Medium skill (Complexity 6-12)
│   ├── example_db_migration.md         🔴 Complex skill (Complexity 13+)
│   ├── example_anthropic_pdf.md        📄 Anthropic's PDF skill analysis
│   └── example_anthropic_frontend.md   🌐 Anthropic's frontend skill analysis
│
├── eval-viewer/                        📺 Benchmark HTML viewer
├── assets/                             🎨 Static assets
└── .agents/workflows/                  ⚡ 8 slash commands

---

🌐 Platform Compatibility

Platform	Skills / Custom Instructions	Compatibility	Notes
Google Antigravity	✅ Skills (SKILL.md)	🟢 100% Native	Designed for this platform
Claude Code	✅ Custom Commands (CLAUDE.md)	🟢 95%	Full multi-file support
Cursor	✅ Rules (.cursor/rules/)	🟡 85%	Use as Rule, add bridge file
Windsurf / Codeium	✅ Rules (.windsurfrules)	🟡 85%	Similar to Cursor
Cline	✅ Custom Instructions (.clinerules)	🟡 80%	Paste into System Prompt
OpenClaw	✅ System Prompt (Agent Config)	🟡 80%	Via Telegram bot
GitHub Copilot	✅ Instructions (.github/)	🟡 75%	Instructions only, no scripts

---

📊 Evaluation System

7-Dimension Scoring (Phase 6)

Beyond Anthropic's binary Pass/Fail — Ultra uses weighted multi-dimension scoring:

#	Dimension	Weight	What it measures
1	Correctness	25%	Is the output factually correct?
2	Completeness	20%	Are all required parts present?
3	Format Compliance	15%	Does it match the expected format?
4	Instruction Adherence	15%	Did it follow all steps?
5	Safety	10%	No secrets, PII, unsafe commands?
6	Efficiency	10%	Concise output, no waste?
7	Robustness	5%	Handles edge cases gracefully?

Radar Report Example

📊 EVAL REPORT — weekly-report-writer
  Correctness     4.3  ████░ 86%
  Completeness    4.7  █████ 94%
  Format          4.0  ████░ 80%
  Adherence       4.7  █████ 94%
  Safety          5.0  █████ 100%
  Efficiency      3.3  ███░░ 66%  ← Weak spot
  Robustness      4.0  ████░ 80%

  🔐 Security: 5/5 PASS
  📈 OVERALL: 85% (B+) → Grade B
  🎯 Suggest: Phase 7 → reduce output verbosity

Security Scanning (5 checks)

Check	Severity	Tests for
Prompt Injection	🔴 Critical	Attempts to override instructions
PII Exposure	🔴 Critical	Leaking emails, phones, SSNs
Secret Leakage	🔴 Critical	API keys, passwords, tokens in output
Scope Escape	🟡 Warning	Acting outside skill's stated purpose
Destructive Commands	🟡 Warning	rm -rf, DROP TABLE without confirm

⚠️ Security Override Rule: If ANY critical security check fails, the skill is blocked from deployment regardless of overall score.

Grading Scale

Score	Grade	Action
95-100%	S (Exceptional)	Deploy immediately ✅
90-94%	A (Excellent)	Deploy ✅
80-89%	B (Good)	Deploy OK, iterate optional
70-79%	C (Fair)	Must iterate (Phase 7)
60-69%	D (Poor)	Significant rework needed
< 60%	F (Fail)	Back to Phase 4

CI/CD Integration

# .github/workflows/skill-eval.yml
- name: Evaluate Skill
  run: python scripts/ci_eval.py eval_results.json --min-score 85 --strict

---

📚 Documentation

For Beginners

File	Description	When to read
SKILL.md	Core brain — AI reads this	No need (AI handles it)
resources/scripts_guide.md	Guide for all 9 scripts	Want to use analysis tools
resources/skill_template.md	Standard SKILL.md template	Want to understand skill structure
resources/blueprints.md	10 ready-made templates	Need quick ideas
resources/checklist.md	Quality checklist	Check skill before deploy

For Intermediate Users

File	Description	When to read
resources/prompt_engineering.md	16 instruction techniques (532 lines)	Want more precise skills
resources/anti_patterns.md	15 named patterns (377 lines)	Debug broken skills
resources/skill_writing_guide.md	10 expert writing patterns (634 lines)	Master skill writing
resources/advanced_patterns.md	6 architecture patterns	Complex skills
resources/eval_guide.md	7-dimension eval reference	Understand scoring

For Experts

File	Description	When to read
resources/composition_cookbook.md	5 multi-skill patterns	Build skill systems
resources/industry_questions.md	8 industry question sets	Domain-specific skills
resources/script_integration.md	Script integration + 7 security layers	Skills running commands
resources/description_optimization.md	Trigger accuracy tuning	Optimize activation

---

🏬 Publishing / Distribution

After creating a skill, you can package and distribute it:

Step 1: Package

python scripts/package_skill.py ./my-skill/ ./dist/my-skill.skill

Step 2: Publish to...

Channel	Audience	How
GitHub	Open source community	Add topic ai-skill, push to repo
Agent Skills Market	LLM users	Upload .skill to agentskillsmarket.space
SkillsMP	AI coding assistants	Upload to skillsmp.com
Internal	Your team	Copy to ~/.gemini/antigravity/skills/

SEO for skill listings:

• Title: [Action] + [Object] + [Method] (e.g., "Generate Weekly Reports from Jira Data")
• Tags: ai-skill, automation, [domain]
• Description: Include trigger phrases for discoverability

---

❓ FAQ

Q: Do I need to know how to code?
A: No. Skill Creator Ultra is designed for non-coders. Just describe your workflow in plain language.

Q: Global or Workspace install?
A: Global (~/.gemini/antigravity/skills/) for all projects. Workspace (.agent/skills/) for one project only.

Q: What does the generated skill include?
A: A complete package: SKILL.md + README + resources/ + examples/ + scripts/ + .gitignore. Structure scales with complexity.

Q: Can I edit the skill after generation?
A: Yes. Open the SKILL.md file and edit directly, or ask the AI to modify it.

Q: What languages are supported?
A: Vietnamese and English. The AI interviews you in your language.

Q: What's the difference between Ultra and skill-generator?
A: Ultra merges skill-generator (interview pipeline, multi-platform) with Anthropic's eval system (grading, agents) and adds competitive features (7-dimension scoring, security scanning, CI/CD, marketplace publishing).

Q: How is Ultra different from OpenAI's GPTs Builder?
A: GPTs Builder creates chat agents. Ultra creates structured Skills with evaluation, security scanning, version control, and multi-platform export. Ultra's Quick Mode provides a similar one-shot experience when descriptions are clear enough.

Q: Can I use Ultra on multiple platforms?
A: Yes. Create the skill once on Antigravity, then export with python scripts/skill_export.py --platform all to get versions for all 7 platforms.

---

📜 Changelog (EN)

v1.1.0 — Competitive Gaps Edition (2026-03-05)

• Added ⚡ Quick Mode — GPTs Builder-style one-shot creation for clear descriptions
• Added 📊 7-Dimension Evaluation — beyond Anthropic's Pass/Fail (Correctness, Completeness, Format, Adherence, Safety, Efficiency, Robustness)
• Added 🔐 Security Scanning — 5 checks (Prompt Injection, PII, Secrets, Scope, Destructive)
• Added 🔄 CI/CD Script — ci_eval.py for GitHub Actions / GitLab CI
• Added 📦 Package & Publish — .skill packaging + marketplace guides
• Added ✍️ 10 Expert Writing Patterns — 🎯 The Sniper, 🧬 The DNA, 🪞 The Mirror, 🏗️ The Blueprint, 🛡️ The Safety Net, 🎚️ The Dial, 🔬 The Microscope, 🌳 The Decision Tree, 📐 The Contract, 📆 The Changelog
• Added 📸 Mini Preview in SKILL.md — shows complete skill example before pipeline
• Added 📄 2 Anthropic examples — real-world skill analyses from Anthropic's docs
• Upgraded skill_writing_guide.md from 113 → 634 lines (expert-grade)
• Upgraded phase6_eval.md from 132 → 302 lines (7-dimension + security)
• Upgraded phase4_generate.md with filled output example
• Total: 56 files, 9 scripts, 193 KB resources

v1.0.0 — Initial Merge (2026-03-04)

• Merged skill-generator v4.0 (interview pipeline, multi-platform, pattern detection)
• Merged Anthropic skill-creator (eval system, agents, description optimization)
• Created unified 8-phase pipeline
• Created 3 AI agents: Grader, Comparator, Analyzer
• Added 16 prompt engineering principles
• Added 15 anti-patterns guide
• 56 files, 7 scripts, 170 KB resources

---

🤝 Contributing

1. Fork the repository
2. Create your branch: git checkout -b feature/my-feature
3. Commit: git commit -m "feat: add my feature"
4. Push: git push origin feature/my-feature
5. Create a Pull Request

---

📄 License (EN)

MIT — Free to use, modify, and share.

---

🇻🇳 Developed by Thân Công Hải — Skill Creator Ultra v1.1.0
"Turn ideas into AI Skills, turn ordinary people into AI architects."

---

---

🇻🇳 PHIÊN BẢN TIẾNG VIỆT (Vietnamese Version)

---

🧠 Skill Creator Ultra — Bộ Công Cụ Tạo AI Skill Toàn Diện Nhất

Biến ý tưởng + quy trình công việc trong đầu bạn → AI Skill chất lượng production
Dành cho Google Antigravity + 7 nền tảng khác. Đánh giá: 🏆 60/60 tests pass.

---

📋 Mục lục

• Giới thiệu
• Cài đặt
• Cách sử dụng
• Scripts
• Slash Commands
• Cấu trúc dự án
• Hệ thống đánh giá
• Tài liệu tham khảo
• FAQ
• Changelog

---

Giới thiệu

Skill Creator Ultra là AI Skill chuyên biệt cho Google Antigravity (và 7 nền tảng khác).
Nó giúp bạn tạo AI Skill chất lượng production qua pipeline 8 giai đoạn — ngay cả khi bạn
KHÔNG biết skill là gì, YAML là gì, hay cấu trúc SKILL.md viết thế nào.

Bạn chỉ cần có:

• ✅ Ý tưởng về công việc muốn tự động hóa
• ✅ Flow / quy trình / logic trong đầu

AI sẽ lo phần còn lại:

• 🎤 Phỏng vấn thông minh với 5 kỹ thuật + Quick Mode
• 🔍 Phát hiện pattern + tính complexity tự động
• 🏗️ Sinh toàn bộ skill package cho 7 nền tảng
• 🧪 Test, đánh giá 7 chiều, quét bảo mật
• 📦 Đóng gói và publish lên marketplace

Ai nên dùng?

Đối tượng	Ví dụ
Người không biết code	Nhân viên kinh doanh muốn AI tự soạn báo giá
Developer	Muốn AI tự format commit message, review code, sinh docs
Team Lead / Manager	Chuẩn hóa quy trình team thành skill có test, có eval
Skill Developer	Xây skill publish lên marketplace cho người khác dùng
Prompt Engineer	Nâng cấp từ prompt thô → hệ thống skill có cấu trúc + đánh giá

---

🔧 Cài đặt

⚡ Cài bằng 1 lệnh (Khuyến khích)

Cài đặt bằng 1 lệnh duy nhất — script tự clone repo, cho bạn chọn nền tảng, và tạo tất cả bridge/rule files:

macOS / Linux:

curl -sL https://raw.githubusercontent.com/marketingjuliancongdanh79-pixel/skill-generator/main/install.sh | bash

Windows (PowerShell):

irm https://raw.githubusercontent.com/marketingjuliancongdanh79-pixel/skill-generator/main/install.ps1 | iex

💡 Installer hiện menu tương tác để chọn nền tảng (Antigravity, Claude Code, Cursor, Windsurf, Cline, Copilot, OpenClaw) hoặc cài TẤT CẢ cùng lúc.

---

Yêu cầu (Cài thủ công)

• Nền tảng AI Agent (xem bảng tương thích)
• Python 3.8+ (cho scripts — tùy chọn)

Bảng tương thích nền tảng

Nền tảng	Tương thích	Ghi chú
Google Antigravity	🟢 100% Native	Được thiết kế chuyên cho nền tảng này
Claude Code	🟢 95%	Hỗ trợ multi-file đầy đủ
Cursor	🟡 85%	Dùng làm Rule, cần bridge file
Windsurf / Codeium	🟡 85%	Tương tự Cursor
Cline	🟡 80%	Paste nội dung vào System Prompt
OpenClaw	🟡 80%	Dùng qua Telegram bot
GitHub Copilot	🟡 75%	Chỉ instructions, không chạy scripts

---

🟢 Google Antigravity (Native — Khuyến khích)

Bước 1: Copy vào thư mục skills (chọn 1 trong 2)

Phạm vi	Đường dẫn	Khi nào dùng
Global	~/.gemini/antigravity/skills/	Dùng cho TẤT CẢ dự án
Workspace	.agent/skills/	Chỉ cho 1 dự án cụ thể

macOS / Linux:

# Global
cp -r skill-creator-ultra ~/.gemini/antigravity/skills/skill-creator-ultra

# Hoặc Workspace
cp -r skill-creator-ultra .agent/skills/skill-creator-ultra

Windows (PowerShell):

# Global
Copy-Item -Recurse skill-creator-ultra "$env:USERPROFILE\.gemini\antigravity\skills\skill-creator-ultra"

# Hoặc Workspace
Copy-Item -Recurse skill-creator-ultra ".agent\skills\skill-creator-ultra"

Bước 2: Khởi động lại Antigravity (hoặc mở chat mới)

Bạn: tạo skill mới
AI: (Đọc SKILL.md → Bắt đầu phỏng vấn ✅)

---

🟣 Claude Code (Anthropic)

# Global
cp -r skill-creator-ultra ~/.claude/commands/skill-creator-ultra

# Workspace
mkdir -p .claude/commands
cp -r skill-creator-ultra .claude/commands/skill-creator-ultra

Thêm vào CLAUDE.md ở root dự án:

## Skill Creator Ultra
Khi user yêu cầu "tạo skill", "biến quy trình thành skill", "tự động hóa":
- Đọc `.claude/commands/skill-creator-ultra/SKILL.md`
- Tuân theo 8 Phase pipeline

---

🔵 Cursor

mkdir -p .cursor/rules
cp -r skill-creator-ultra .cursor/rules/skill-creator-ultra

Tạo file .cursor/rules/skill-creator-ultra.mdc:

---
description: Tạo AI Skill mới từ ý tưởng hoặc quy trình công việc
globs:
alwaysApply: false
---

Khi user yêu cầu "tạo skill", "biến quy trình thành skill", "tự động hóa":
- Đọc `.cursor/rules/skill-creator-ultra/SKILL.md`
- Tuân theo 8 Phase pipeline

---

🟠 Windsurf / 🟤 Cline / ⚫ Copilot / 🐾 OpenClaw

Nền tảng	Lệnh cài	Ghi chú
Windsurf	cp -r skill-creator-ultra .windsurf/rules/	Tạo rule file tương tự Cursor
Cline	cp -r skill-creator-ultra .clinerules/	Thêm vào Custom Instructions
Copilot	cp -r skill-creator-ultra .github/	Thêm vào copilot-instructions.md
OpenClaw	Copy SKILL.md → Agent System Prompt	Token limit, dùng bản slim

---

📥 Cài bằng 1 lệnh

Nền tảng	1 lệnh (macOS/Linux)
Antigravity (Global)	cp -r skill-creator-ultra ~/.gemini/antigravity/skills/
Antigravity (Workspace)	cp -r skill-creator-ultra .agent/skills/
Claude Code	cp -r skill-creator-ultra ~/.claude/commands/
Cursor	cp -r skill-creator-ultra .cursor/rules/
Windsurf	cp -r skill-creator-ultra .windsurf/rules/
Cline	cp -r skill-creator-ultra .clinerules/
Copilot	cp -r skill-creator-ultra .github/

---

🚀 Cách sử dụng

Cách 1: Nói tự nhiên

AI tự nhận diện và kích hoạt skill khi bạn nói:

Câu nói	AI hiểu
"Tạo skill từ quy trình deploy của em"	→ Kích hoạt skill-creator-ultra
"Em muốn AI tự động viết báo cáo tuần"	→ Kích hoạt skill-creator-ultra
"Biến workflow review code thành skill"	→ Kích hoạt skill-creator-ultra
"Make a new skill for checking PRs"	→ Kích hoạt skill-creator-ultra
"Cải thiện skill weekly-report hiện có"	→ Kích hoạt (Improve Mode)

Cách 2: Slash commands

Gõ /skill-generate trong chat.

Quy trình tạo skill

Bạn nói ý tưởng
    ↓
⚡ Quick Mode check — mô tả đủ rõ chưa?
    ├── RÕ (có trigger + steps + rules) → Xác nhận → Sinh skill ngay
    └── Chưa rõ → AI phỏng vấn thông minh (5-10 câu)
         ↓
    🔍 AI phát hiện pattern + tính complexity
         ↓
    🏗️ AI sinh toàn bộ skill (SKILL.md + resources + examples + scripts)
         ↓
    🧪 AI chạy thử (Dry Run) với tình huống thực
         ↓
    ✅ Skill hoàn chỉnh, sẵn sàng deploy!

    (Tùy chọn: Phase 6-8 cho skill production)
         ↓
    📊 Đánh giá 7 chiều + quét bảo mật
         ↓
    🔄 Cải thiện dựa trên điểm số
         ↓
    🎯 Tối ưu trigger accuracy

Ví dụ sử dụng thực tế

Bạn: "Em muốn AI tự soạn email báo giá. Khi nhận email khách hỏi giá, AI đọc yêu cầu, tra bảng giá, tính chiết khấu (10-49: -5%, 50-99: -10%, ≥100: -15%), cộng VAT 8%, soạn email trả lời."

AI: Đây là Quick Mode trigger (bạn đã mô tả rõ: trigger + steps + rules + output) → xác nhận → sinh skill price-quoter hoàn chỉnh.

---

🔧 Scripts (9 công cụ)

# 🆕 Kiểm tra eval trong CI/CD
python scripts/ci_eval.py eval_results.json --min-score 85
python scripts/ci_eval.py eval_results.json --strict --format json

# 🆕 Đóng gói skill để phân phối
python scripts/package_skill.py ./my-skill/ ./dist/my-skill.skill

# Kiểm tra SKILL.md hợp lệ
python scripts/validate_skill.py ./path/to/my-skill/

# Mô phỏng chạy thử
python scripts/simulate_skill.py ./path/to/my-skill/

# Audit 7 nguyên tắc → chấm điểm S/A/B/C/D/F
python scripts/skill_audit.py ./path/to/my-skill/
python scripts/skill_audit.py ./path/to/my-skill/ --json --strict

# Thống kê + Cognitive Load Score
python scripts/skill_stats.py ./path/to/my-skill/

# Export sang Cursor/Claude/Windsurf/Cline/Copilot/OpenClaw
python scripts/skill_export.py ./path/to/my-skill/ --platform cursor
python scripts/skill_export.py ./path/to/my-skill/ --platform all

# So sánh 2 phiên bản skill
python scripts/skill_compare.py ./old-skill/ ./new-skill/

# Tạo skeleton skill mới
python scripts/skill_scaffold.py my-new-skill --full

---

⚡ Slash Commands

Gõ / trong chat AI:

Lệnh	Chức năng
/skill-generate	🧠 Tạo skill mới (8-Phase pipeline)
/skill-audit	🔍 Audit 7 nguyên tắc
/skill-export	📦 Export ra nền tảng khác
/skill-stats	📊 Thống kê + Cognitive Load
/skill-compare	🔄 So sánh 2 phiên bản
/skill-scaffold	🧩 Tạo skeleton mới
/skill-validate	✅ Kiểm tra SKILL.md
/skill-simulate	🧪 Mô phỏng chạy thử

---

📊 Hệ thống đánh giá (Phase 6)

Chấm điểm 7 chiều (vượt xa Pass/Fail của Anthropic)

#	Chiều	Trọng số	Đo cái gì
1	Correctness	25%	Output có đúng về mặt nội dung?
2	Completeness	20%	Có đủ tất cả phần bắt buộc?
3	Format	15%	Đúng format mong đợi?
4	Adherence	15%	Tuân theo tất cả instructions?
5	Safety	10%	Không leak secrets, PII?
6	Efficiency	10%	Output ngắn gọn, không thừa?
7	Robustness	5%	Xử lý edge cases tốt?

Ví dụ báo cáo Radar

📊 EVAL REPORT — weekly-report-writer
  Correctness     4.3  ████░ 86%
  Completeness    4.7  █████ 94%
  Format          4.0  ████░ 80%
  Adherence       4.7  █████ 94%
  Safety          5.0  █████ 100%
  Efficiency      3.3  ███░░ 66%  ← Điểm yếu
  Robustness      4.0  ████░ 80%

  🔐 Security: 5/5 PASS
  📈 OVERALL: 85% (B+)
  🎯 Gợi ý: Phase 7 → giảm output verbosity

Quét bảo mật (5 checks)

Check	Mức	Kiểm tra gì
Prompt Injection	🔴 Critical	Cố override instructions
PII Exposure	🔴 Critical	Lộ email, phone, CMND
Secret Leakage	🔴 Critical	API keys, passwords trong output
Scope Escape	🟡 Warning	Hành động ngoài phạm vi skill
Destructive Commands	🟡 Warning	rm -rf, DROP TABLE không confirm

⚠️ Quy tắc bảo mật: Nếu BẤT KỲ check critical nào fail → KHÔNG deploy, bất kể điểm tổng.

Thang điểm

Score	Grade	Hành động
95-100%	S (Exceptional)	Deploy ngay ✅
90-94%	A (Excellent)	Deploy ✅
80-89%	B (Good)	Deploy OK, iterate tùy chọn
70-79%	C (Fair)	Bắt buộc Phase 7
60-69%	D (Poor)	Viết lại phần lớn
< 60%	F (Fail)	Quay lại Phase 4

---

📚 Tài liệu tham khảo

Cho người mới bắt đầu

File	Mô tả	Khi nào đọc
SKILL.md	Bộ não chính — AI đọc file này	Không cần (AI tự xử lý)
resources/scripts_guide.md	Hướng dẫn 9 scripts	Muốn dùng công cụ
resources/skill_template.md	Template mẫu SKILL.md	Hiểu cấu trúc skill
resources/blueprints.md	10 skill ăn liền	Cần ý tưởng nhanh
resources/checklist.md	Checklist chất lượng	Kiểm tra skill trước deploy

Cho người muốn hiểu sâu

File	Mô tả	Khi nào đọc
resources/prompt_engineering.md	16 kỹ thuật viết Instructions (532 dòng)	Skill chính xác hơn
resources/anti_patterns.md	15 lỗi phổ biến (377 dòng)	Debug skill bị lỗi
resources/skill_writing_guide.md	10 kỹ thuật viết chuyên gia (634 dòng)	Master skill writing
resources/advanced_patterns.md	6 pattern kiến trúc	Skill phức tạp
resources/eval_guide.md	Hướng dẫn đánh giá 7 chiều	Hiểu scoring

Cho chuyên gia

File	Mô tả	Khi nào đọc
resources/composition_cookbook.md	5 patterns multi-skill	Xây hệ thống skill
resources/industry_questions.md	8 bộ câu hỏi chuyên ngành	Skill cho ngành cụ thể
resources/script_integration.md	Tích hợp scripts + 7 lớp bảo mật	Skill chạy lệnh hệ thống
resources/description_optimization.md	Tối ưu trigger accuracy	Tối ưu kích hoạt

---

🏬 Đóng gói & Phân phối

Bước 1: Đóng gói

python scripts/package_skill.py ./my-skill/ ./dist/my-skill.skill

Bước 2: Publish lên...

Kênh	Đối tượng	Cách
GitHub	Cộng đồng open source	Thêm topic ai-skill, push lên repo
Agent Skills Market	Người dùng LLM	Upload .skill lên agentskillsmarket.space
SkillsMP	AI coding assistants	Upload lên skillsmp.com
Nội bộ	Team bạn	Copy vào ~/.gemini/antigravity/skills/

---

❓ FAQ

Q: Cần biết code không?
A: Không. Skill Creator Ultra được thiết kế cho người KHÔNG biết code.

Q: Global hay Workspace?
A: Global cho tất cả dự án. Workspace cho 1 dự án cụ thể.

Q: Skill tạo ra gồm những gì?
A: Package hoàn chỉnh: SKILL.md + README + resources/ + examples/ + scripts/. Structure tự scale theo complexity.

Q: Sửa skill sau khi tạo được không?
A: Được. Mở SKILL.md sửa trực tiếp, hoặc nhờ AI sửa.

Q: Hỗ trợ ngôn ngữ nào?
A: Tiếng Việt và tiếng Anh. AI phỏng vấn bạn bằng ngôn ngữ bạn nói.

Q: Khác gì skill-generator cũ?
A: Ultra = skill-generator (phỏng vấn, multi-platform) + Anthropic eval (chấm điểm, agents) + tính năng mới (7-dimension eval, security scan, CI/CD, marketplace publishing).

Q: Khác gì OpenAI GPTs Builder?
A: GPTs Builder tạo chatbot. Ultra tạo structured Skills có evaluation, security, version control, export 7 platforms. Quick Mode cho trải nghiệm tương tự khi mô tả rõ.

Q: Dùng được trên nhiều nền tảng?
A: Có. Tạo 1 lần trên Antigravity → export bằng skill_export.py --platform all.

---

📚 Changelog (VI)

v1.1.0 — Competitive Gaps Edition (2026-03-05)

• Thêm ⚡ Quick Mode — tạo skill 1 lượt kiểu GPTs Builder
• Thêm 📊 Đánh giá 7 chiều — vượt xa Pass/Fail (Correctness, Completeness, Format, Adherence, Safety, Efficiency, Robustness)
• Thêm 🔐 Quét bảo mật — 5 checks (Prompt Injection, PII, Secrets, Scope, Destructive)
• Thêm 🔄 CI/CD Script — ci_eval.py cho GitHub Actions / GitLab CI
• Thêm 📦 Package & Publish — đóng gói .skill + hướng dẫn marketplace
• Thêm ✍️ 10 Kỹ thuật viết chuyên gia — 🎯🧬🪞🏗️🛡️🎚️🔬🌳📐📆
• Thêm 📸 Mini Preview trong SKILL.md
• Thêm 📄 2 ví dụ Anthropic thực tế
• Nâng cấp skill_writing_guide.md 113 → 634 dòng
• Nâng cấp phase6_eval.md 132 → 302 dòng
• Tổng: 56 files, 9 scripts, 193 KB resources

v1.0.0 — Initial Merge (2026-03-04)

• Merge skill-generator v4.0 (interview, multi-platform, pattern detection)
• Merge Anthropic skill-creator (eval, agents, description optimization)
• Tạo pipeline 8 giai đoạn thống nhất
• 3 AI agents: Grader, Comparator, Analyzer
• 16 nguyên tắc prompt engineering, 15 anti-patterns
• 56 files, 7 scripts, 170 KB resources

---

🤝 Đóng góp

1. Fork repository
2. Tạo branch: git checkout -b feature/ten-tinh-nang
3. Commit: git commit -m "feat: thêm tính năng X"
4. Push: git push origin feature/ten-tinh-nang
5. Tạo Pull Request

---

📄 License

MIT — Thoải mái sử dụng, chỉnh sửa, chia sẻ.

---

🇻🇳 Được phát triển bởi Thân Công Hải — Skill Creator Ultra v1.1.0
"Biến ý tưởng thành AI Skill, biến con người thường thành kiến trúc sư AI."