Quick Start
You don't write KERN. You tell an AI what you want, it writes KERN, you compile it. That's the workflow.
The workflow
Describe what you want
In plain English to any LLM (Claude, GPT, Gemini, etc.)
LLM writes .kern
The AI generates KERN — a compact, structured format it naturally understands
You compile
kern dev app.kern --target=nextjs and you have production code
Review & ship
kern review catches bugs. You deploy.
1. Install KERN
npm install -g @kernlang/cli2. Prompt your LLM
Open Claude, ChatGPT, or any LLM. Give it the KERN spec and tell it what to build:
You are an expert in KERN, a compiled UI language.
Write a .kern file for a SaaS landing page with:
- Dark theme, orange accents
- Hero section with headline and two CTA buttons
- 3-column feature grid
- Pricing section with Free and Pro tiers
- Footer with links
Use the KERN spec from https://kernlang.dev/llm/spec.kernThe LLM generates a complete .kern file. You don't need to know the syntax — the AI does.
3. Save and compile
Paste the LLM output into a file and compile:
# Save the LLM output
pbpaste > landing.kern
# Compile to Next.js (watches for changes)
kern dev landing.kern --target=nextjs --outdir=app/
# Or compile to Vue
kern dev landing.kern --target=vue --outdir=src/
# Or compile to Express API
kern dev api.kern --target=express --outdir=src/4. Review the output
Before shipping, run kern review on the generated code to catch bugs the AI might have introduced:
kern review app/ --recursive5. Iterate with the LLM
Want changes? Just tell the AI:
"Add a testimonials section between features and pricing.
Make the hero headline bigger. Add a dark/light mode toggle."The AI updates the .kern file. You recompile. That's it.
Why not just ask the AI to write React directly?
- 70% fewer tokens — KERN is ~3x more compact. Cheaper, faster, less hallucination.
- 11 targets from 1 source — Same .kern compiles to Next.js, Vue, Express, Python, Native, etc.
- Guaranteed valid output — The compiler catches structural errors. Raw LLM code often has subtle bugs.
- Review built in — 68 rules catch security issues, dead logic, and framework mistakes automatically.