LLM Integration

LightShell is designed from the ground up so that AI agents can generate working desktop apps. The entire API surface is small, consistent, and discoverable — there are no hidden configuration files, no complex build steps, and no framework-specific abstractions. An AI that knows the 15 lightshell.* namespaces can build a functional desktop app in a single prompt.

This section covers the three integration points that make this work.

Why LightShell Works Well with AI

Most desktop frameworks have steep learning curves — complex build systems, multiple process models, and large API surfaces. LightShell keeps things simple:

• One global object. Every API lives under window.lightshell.* — no imports, no modules, no bundlers.
• Every call is async. No callback patterns, no sync/async split. Always await lightshell.something().
• Plain HTML/CSS/JS. The entry point is src/index.html. No JSX, no templates, no compilation step.
• AI-friendly errors. Permission denials and API errors include the exact fix needed, so an AI can self-correct.

Integration Points

1. llms.txt — Context Injection

The fastest way to give an AI knowledge of LightShell. Two files are available:

• llms.txt — a compact summary (~30 lines) with links to full documentation. Best for quick context in chat-based tools.
• llms-full.txt — the complete API reference (~800 lines). Best for comprehensive code generation where the AI needs every method signature and parameter.

Paste the URL into your AI conversation or add it to a system prompt.

2. Prompting Guide — How to Ask

Not all prompts are equal. The prompting guide covers how to structure requests so an AI produces working LightShell apps on the first try. Includes a template, tips, and guidance on what works well versus what needs iteration.

3. IDE Rules — Cursor and AGENTS.md

For AI-assisted editors, project-level rule files prevent common mistakes before they happen:

• Cursor Rules — a .cursorrules file that teaches Cursor about LightShell APIs, constraints, and platform differences.
• AGENTS.md — a project-level instruction file for Claude Code, Devin, and other agent-based tools.

Quick Start

The fastest path from zero to a working app:

1. Open your AI tool (Claude, ChatGPT, Cursor, Copilot).
2. Paste this context URL: https://lightshell.dev/llms-full.txt
3. Use one of the example prompts or write your own.
4. Save the generated files into a LightShell project (lightshell init my-app).
5. Run lightshell dev to see it live.

If the AI makes a mistake, check the common mistakes page for the fix — most issues are the same patterns repeated.

All Pages in This Section

Page	What It Covers
llms.txt Specification	The llms.txt and llms-full.txt files for AI context injection
Prompting Guide	How to structure prompts for best results
Cursor Rules	.cursorrules file for Cursor IDE
AGENTS.md	Project-level instructions for Claude Code and other agents
Example Prompts	Ready-to-use prompts for common app types
Common AI Mistakes	Frequent errors and their fixes