Autonomously create research-grounded, high-quality Claude Code skills.

Unlike template-based skill creation, Skill Forge researches official documentation and domain sources before generating any content. Every fact is verified, every procedure is sourced.

• Research-First Architecture — Fetches Claude Code docs and domain docs before generation
• Source Verification — Every fact traced to official documentation
• Citation Included — sources.md with all URLs and retrieval dates
• Official Format — Follows latest Anthropic skill patterns
• Progressive Disclosure — Lean SKILL.md with detailed references/

claude --add-plugin https://github.com/VAMFI/skill-forge

Via Git Clone:

git clone https://github.com/VAMFI/skill-forge ~/.claude/plugins/skill-forge

Manual Download:

# Download and extract to plugins directory curl -L https://github.com/VAMFI/skill-forge/archive/main.tar.gz | tar -xz -C ~/.claude/plugins/ mv ~/.claude/plugins/skill-forge-main ~/.claude/plugins/skill-forge

Per-Session Usage:

claude --plugin-dir /path/to/skill-forge

/skill-forge:forge-skill Create a skill for Kubernetes deployment best practices /skill-forge:forge-skill Build a skill for Python pytest testing /skill-forge:forge-skill I need a skill for PostgreSQL optimization --output ./my-plugin/skills

1. Research Phase: Agent fetches:

   • Claude Code skill documentation (DeepWiki, WebFetch)
   • Domain official documentation (WebSearch, WebFetch, Context7)

2. Design Phase: Agent plans:

   • Trigger phrases for description
   • SKILL.md structure
   • Reference files needed

3. Create Phase: Agent generates:

   • SKILL.md with proper frontmatter
   • references/ with detailed content
   • sources.md with citations

4. Validate Phase: Agent checks:

   • Structure completeness
   • Description quality
   • Source documentation

The main autonomous agent that:

• Researches documentation
• Designs skill structure
• Creates all files
• Validates output

Tools Used:

• WebSearch, WebFetch (web research)
• DeepWiki (GitHub repo analysis)
• Context7 (library documentation)
• Read, Write, Bash (file operations)
• TodoWrite (progress tracking)

Core knowledge for skill creation:

• Official SKILL.md format
• Research methodology
• Quality checklist
• Source hierarchy

Entry point for skill creation with argument support.

The agent consults these authoritative sources:

Plus domain-specific official documentation via WebSearch/WebFetch/Context7.

The agent prioritizes sources in this order:

1. Official documentation (highest trust)
2. Official GitHub repos
3. Authoritative tech sources (MDN, framework docs)
4. Verified community (high-vote Stack Overflow)
5. General web (marked as unverified)

Created skills follow this structure:

skill-name/ ├── SKILL.md # Core instructions (≤2,000 words) ├── references/ │ ├── detailed-guide.md # Extended content │ ├── patterns.md # Common patterns │ └── sources.md # All citations └── examples/ # Working code (if applicable) └── example.sh 

User: /skill-forge:forge-skill Create a skill for Kubernetes deployment Agent: I'll create a research-grounded skill for Kubernetes deployment. Let me research the official documentation first... [Agent fetches Claude Code skill format from DeepWiki] [Agent fetches Kubernetes official docs] [Agent researches K8s deployment best practices] Agent: Where should I create this skill? 1. ./skills/kubernetes-deployment/ 2. Custom path User: Option 1 [Agent creates skill with full citations] Agent: ✓ Skill created at ./skills/kubernetes-deployment/ Files: - SKILL.md (1,847 words) - references/deployment-patterns.md - references/kubectl-reference.md - references/sources.md (12 citations) Sources consulted: - kubernetes.io/docs (official) - anthropics/skills (official) - CNCF best practices guide 

• Skills are grounded in official documentation
• No hardcoded credentials
• Always review generated skills before production use
• The agent follows Anthropic's security guidelines

• Claude Code with plugin support
• Internet access for documentation fetching
• MCP servers: DeepWiki, Context7 (optional but recommended)

MIT

Contributions welcome! Please ensure any changes maintain the research-first philosophy.