The meta-skill for creating, importing, and auditing agent skills.
Create · Import · Audit · Secure · Ship
Built primarily on The Complete Guide to Building Skills for Claude by Anthropic, with security hardening from real-world threat research.
The agent skills ecosystem is growing fast. SkillHub alone hosts 24,000+ skills. Snyk's ToxicSkills research found that 13% of community skills contain critical security flaws, including credential theft, prompt injection, and hidden malware. Creating a well-structured skill from scratch means knowing the spec, best practices, and dozens of small decisions.
skill-master solves both problems:
- Creating skills — Interactive guided workflow that explains every decision in plain English, generates spec-compliant output, and recommends complementary skills from the ecosystem.
- Importing skills — Automated security scanning before installation. Every skill gets checked against 40+ threat patterns derived from real malware samples.
- Reviewing skills — Quality assessment against Anthropic's official best practices with actionable improvement suggestions.
git clone https://github.com/hexcantcode/skill-master.git ~/.claude/skills/skill-masterskill-master activates automatically when you ask Claude to create, import, or review a skill. You can also invoke it directly:
/skill-master
Create a skill:
"Help me create a skill for reviewing pull requests"
Import a skill from a URL:
"Import this skill: https://github.com/trailofbits/skills"
Review an existing skill:
"Review my solidity-auditor skill for quality and security"
skill-master/
├── SKILL.md # Main entry — mode detection + routing
│
├── references/
│ ├── creation-workflow.md # 7-step interactive creation guide
│ ├── import-audit-workflow.md # Fetch → Scan → Gate → Install
│ ├── security-checklist.md # 5 threat categories, 40+ patterns
│ ├── quality-checklist.md # Anthropic best practices evaluation
│ ├── spec-quick-reference.md # Agent Skills spec condensed
│ └── supplementary-skills-guide.md # How to find complementary skills
│
├── scripts/
│ ├── validate_skill.py # Structure + frontmatter validator
│ └── security_scan.py # ToxicSkills-based threat scanner
│
├── assets/
│ ├── skill-template.md # Starter SKILL.md template
│ └── github-readme-template.md # README template for hosting skills
│
└── LICENSE # MIT
Walks you through an interactive flow:
- Understand — Asks about purpose, audience, and use cases
- Decide — Presents 4 critical architectural decisions, each explained in plain English for non-technical users:
- Who can trigger it? (auto vs manual)
- Where should it run? (inline vs sandboxed)
- What tools can it access? (unrestricted vs locked down)
- How complex is it? (single file vs full toolkit)
- Draft — Generates spec-compliant SKILL.md with proper frontmatter
- Supplement — Recommends complementary skills from SkillHub, Anthropic's official repo, and your installed skills
- Validate — Runs automated structure and security checks
- Scaffold — Creates the full directory
- Ship — Optionally generates a GitHub README and guides you through publishing
Security scan → Gate → Validate → Quality check → Supplement → Install
Every imported skill passes through the security scanner first. The scanner produces one of three ratings:
| Rating | Score | Action |
|---|---|---|
| SAFE | 0-1 | Proceed with installation |
| CAUTION | 2-5 | Each finding explained in plain English, requires your explicit approval |
| DANGER | 6+ | Installation blocked, with explanation of what was found |
Combines both scanners on a skill you already have installed, plus a manual quality assessment against Anthropic's best practices checklist. Outputs a rating (EXCELLENT / GOOD / NEEDS WORK / POOR) with specific fixes.
The security scanner checks for 5 threat categories based on Snyk's ToxicSkills study of 3,984 community skills:
| Category | What It Detects | Real-World Prevalence |
|---|---|---|
| Prompt Injection | Role overrides, safety suppression, instruction hijacking | 91% of malicious skills |
| Malicious Code | Download-and-execute, obfuscated payloads, privilege escalation | 100% of confirmed malware |
| Credential Theft | Env var harvesting, SSH/AWS/Docker credential access | 63% of malicious skills |
| Security Disablement | Firewall disabling, history clearing, shell profile backdoors | Common in destructive skills |
| Data Exfiltration | Outbound POST with local data, DNS exfiltration, encoded transmission | 54% use third-party fetching |
The scanner is code-fence aware — it won't flag patterns that appear inside markdown code blocks (documentation examples), only actual instructions and executable code.
# Security scan
python3 ~/.claude/skills/skill-master/scripts/security_scan.py /path/to/skill
# Structure validation
python3 ~/.claude/skills/skill-master/scripts/validate_skill.py /path/to/skill| Platform | Status |
|---|---|
| Claude Code | Fully supported |
| Codex CLI | Agent Skills compatible |
| Gemini CLI | Agent Skills compatible |
| OpenCode | Agent Skills compatible |
| Cursor | Agent Skills compatible |
The skill follows the Agent Skills open standard, so it works with any compatible agent. The Python scripts require Python 3.8+ (stdlib only, no dependencies).
Built on research and specifications from:
- The Complete Guide to Building Skills for Claude — Anthropic (2026)
- Agent Skills Specification — Open standard
- Skill Authoring Best Practices — Anthropic
- Claude Code Skills Documentation — Anthropic
- ToxicSkills: Agent Skills Supply Chain Compromise — Snyk (2026)
- SkillHub — Community marketplace (24,000+ skills)
MIT — see LICENSE for details.