Skill Fuser is an AI-powered tool that merges and compresses multiple AI Agent SKILL.md files into optimized outputs. Based on the SkillReducer research paper, which found that only 38.5% of Skill content is actionable core rules — and removing the rest actually improves Agent performance by 2.8%.
Fusion 50% — Aggressive Compression
Keeps only core rules, removes examples & background
Fusion 90% — Light Compression
Preserves most details, merges duplicates only
Analysis — Content Audit
Classifies every paragraph by importance, outputs statistics & recommended budget
Upload Skills → Classify by Type → Merge Same-Type Skills → Present All Results
Step 1: AI Classification Each uploaded Skill is classified into one of 30 categories (based on the VoltAgent/awesome-openclaw-skills taxonomy):
| Category | Examples |
|---|---|
| Web & Frontend | React patterns, CSS conventions, accessibility |
| AI & LLMs | Prompt engineering, model selection, API usage |
| Security | OWASP checklists, auth procedures, threat models |
| DevOps & Cloud | CI/CD, deployment, Docker, monitoring |
| Git & GitHub | Branching, PR workflows, commit conventions |
| + 25 more | CLI, Data, Gaming, Health, IoT, Media... |
Step 2: Type-Aware Merging
- Skills of the same category are merged using a category-specific prompt (28 custom prompts total)
- e.g., Security skills merge with "NEVER remove any security rule"
- e.g., Web skills merge with "Unify component patterns, keep the most robust version"
- Skills in categories not suited for merging are kept as-is
- Budget is split equally across mergeable groups
Step 3: Grouped Results Results are presented in collapsible groups:
- Merged groups — show merged output with skill names
- Kept Separate — show original content for standalone skills
Upload Skills → AI classifies every paragraph by importance (Core Rule / Background / Example / Template / Redundant) → outputs statistics report with recommended token budget
- 30 Skill Categories — classification based on VoltAgent/awesome-openclaw-skills taxonomy
- 28 Custom Merge Prompts — each category has a specialized prompt for higher quality merges
- Unknown Category Detection — if AI returns an unrecognized type, it's reported to the user and kept separate
- Token Budget Control — set output limit, budget auto-split across merge groups
- Multi-model Support — OpenAI, Anthropic, Google Gemini, DeepSeek, custom endpoints
- Privacy First — pure frontend, all data stays in your browser, no backend
- History & Favorites — save, search, and revisit past fusion results
- Data Management — export/import all data as JSON backup
Visit: https://skill-fuser.vercel.app — no signup, no install.
- Click "Add API Key" (top right) → select AI provider → enter API Key
- Paste Skill content or upload
.mdfiles in the left panel - Set Target Output token budget
- Click "Start Fusion"
- Review grouped results — copy or download
git clone https://github.com/Thomaszhou22/skill-fuser.git
cd skill-fuser
npm install
npm run dev # Development
npm run build # Production| Provider | Default Model | Notes |
|---|---|---|
| OpenAI | gpt-4o-mini | gpt-4o, gpt-4.1 series |
| Anthropic | claude-sonnet-4 | Sonnet/Haiku |
| Google Gemini | gemini-2.0-flash | Flash/Pro series |
| DeepSeek | deepseek-chat | Auto-configured |
| Custom | — | Any OpenAI-compatible endpoint |
Recommended: Gemini Flash or GPT-4o-mini — fast, cheap, good enough.
Based on SkillReducer: Optimizing LLM Agent Skills for Token Efficiency:
| Finding | Data |
|---|---|
| Core rules in Skill files | Only 38.5% |
| Agent performance after removing non-essential content | +2.8% |
| Max compression ratio (lossless) | 60%+ |
5-Level Classification: Core Rule → Background → Example → Template → Redundant
Compression Pipeline: Classify → Deduplicate → Compress → Progressive Disclosure
React 19 + TypeScript + Tailwind CSS v4 + Vite 6 + Vercel
This project is licensed under the GNU Affero General Public License v3.0.
In simple terms: You can freely use, study, and modify this project. However, if you distribute a modified version (including as a network service), you must also release your source code under AGPL-3.0.
What you can do:
- ✅ Personal use, learning, and research
- ✅ Modify and adapt for any purpose
- ✅ Use as a network service
What you must do:
- 📋 If you distribute or offer it as a network service, release your modified source code under AGPL-3.0
- 📝 Include the original copyright notice and license text
Core principle: Share alike — if you improve it, share your improvements with the community.