Diffuser

Developer Controlled, AI Accelerated Software Creation

Diffuser is the interaction layer between LLMs, humans, and code—keeping control in the hands of the developer by making every change explicit, reversible, and verifiable.

No more copy-pasting entire files. No more losing features when AI regenerates code. No more giving AI autonomy over your code. Diffuser applies AI-generated patches directly into your codebase with surgical precision.

Why Diffuser?

🎯 AI-Powered. Developer-Controlled.

Your current LLM can make code, but it can't change your source files. Diffuser patches allow you to change files by copying the patch generated by the LLM and applying it to your codebase in two clicks.

✨ Complete Features, Every Time

LLMs often omit critical features when regenerating code or stop mid-way. Diffuser uses surgical patches that modify only what needs to change, preserving your existing code and preventing feature drift across iterations.

⏮️ Patch-Level Undo/Redo

Avoid multi-file change chaos. Diffuser tracks each patch as a discrete, reversible action, so you always know exactly what you're undoing or redoing.

🔍 Crystal Clear Context

Never wonder which file or function needs editing. Diffuser patches explicitly specify target paths and exact line matches, eliminating ambiguity and making every change reviewable and traceable.

🔄 Error Feedback Loop

Diffuser integrates seamlessly with VS Code diagnostics, catching issues in real-time and allowing you to feed them back to the LLM for fixes.

💰 Save the Tokens

Stop wasting tokens. Diffuser's patches drastically reduce output tokens, sending only precise changes instead of regenerating entire files every turn.

⚡ Export Your Codebase at the Speed of Light

Pick your files and export them instantly. Diffuser streamlines codebase sharing with flexible export options—send full files or outlines, along with the folder structure, to your AI tools.

🔒 Secure Your Source Code

Never risk exposing sensitive code snippets. Keep your intellectual property secure with outline-only exports. Diffuser lets you share code structure and context without leaking proprietary implementation details.

Support & Community

• Website: diffuser.dev (Made with Diffuser)
• Feedback, Feature Requests, Bugs: diffuser.dev/feedback

How It Works

Diffuser uses an interactive 3D prism interface where each face provides different functionality.

• Navigation: Left mouse button hold and drag to rotate the prism.
• Action: Right-click a face to engage its functionality.

Quick Start: Mastering the Terminus Prism

1. FRONT FACE: ∆ — Apply AI Patches

The core interface for lightning-fast code updates.

• Action: Copy the patch text generated by your AI.
• Execute: On the Front Face, Right-click once to VALIDATE (preview) and Right-click again to APPLY.
• Result: Multi-file changes are applied in milliseconds.

2. TOP FACE: Τ — Train Your AI

Prime your assistant to output patches in Diffuser format.

• Action: Drag down to see the Top Face. Right-click to copy the System Prompt.
• Result: Paste this into your AI (Claude, ChatGPT, etc.) to teach it how to generate surgical patches.

3. LEFT FACE: Ψ — Export Context

Give the AI the "big picture" without burning tokens.

• Action: Drag right to see the Left Face. Right-click files in your tree to mark them.
• Configure: Left-click the file icons to cycle between ⬤ Full (content), ⭘ Outline (structure), or 🔴 Excluded.
• Result: Right-click the Prism to copy your selection and file tree structure to the clipboard.

4. RIGHT FACE: Σ — Error Detection

Close the feedback loop with real-time diagnostics.

• Status: If VS Code detects errors after a patch, this face will flash red.
• Action: Drag left to the Right Face and Right-click to copy the error details.
• Result: Paste the errors back to your AI to get an immediate fix-patch.

5. BOTTOM FACE: ◀ ▶ — Undo / Redo

Time-travel through your changes with total safety.

• Action: Drag up to the Bottom Face.
• Control: Right-click the left half to Undo; Right-click the right half to Redo.
• Result: Reverses the changes from the previous patch.

The Diffuser Format

Diffuser uses a simple, explicit patch format that LLMs can easily generate:

∆NEW∆ src/components/Button.tsx
export const Button = () => {
  return <button>Click me</button>;
};

∆REPLACE∆ src/App.tsx
import React from 'react';
+import { Button } from './components/Button';

∆REPLACE∆ src/App.tsx
  return <div>Hello World</div>;
-  return <div>Hello World</div>;
+  return (
+    <div>
+      Hello World
+      <Button />
+    </div>
+  );

∆DELETE∆ src/old-component.tsx

Operations:

• ∆NEW∆ - Create new files or folders
• ∆REPLACE∆ - Modify existing files (Add/Remove lines)
• ∆DELETE∆ - Remove files or folders

Tips & Best Practices

🤖 The Gold Standard: Use Claude

While Diffuser works with any LLM, Claude Sonnet is the recommended partner for the highest quality patches.

🎯 Use with Git

Diffuser works beautifully with version control, allowing you to easily review what the LLM changed in your code.

📦 Export Smart, Save Tokens

Don’t burn your context window. Use Outline Mode (⭘) for large library files—it gives the AI the "map" without wasting tokens on implementation. Use 🔴 Exclude for folders like node_modules/ or dist/.

🧠 Remind Your AI (Fixing Patch Failures)

If an AI generates a patch that fails, rotate to the Left Face, re-export those specific files as Full (⬤), and paste them back. This "reminds" the AI of the exact line matches.

🐛 Close the Error Loop

When a patch introduces a bug, don’t fix it manually. Rotate to the Right Face, copy the diagnostics, and feed them back to the AI for a follow-up patch.

Workflow Example

Scenario: Add authentication to your React app

1. Export context (Left Face): Select src/App.tsx, src/routes/, and package.json. Right-click the prism to copy.
2. Ask AI: Paste the context and ask: "Add JWT-based auth with login/logout and protected routes."
3. AI responds: AI provides Diffuser patches for auth.tsx, App.tsx, and package.json.
4. Validate (Front Face): Copy the patch, Right-click the prism → VALIDATE.
5. Apply (Front Face): Right-click again → APPLY.
6. Test & iterate: If errors appear, copy them from the Right Face and send them back to the AI.

License

Proprietary freeware - Free to use (including commercially). Software cannot be resold, repackaged, or included in any other applications. Source code not available for the public.

All rights reserved by PROJECT500.

**Made with 💜 by PROJECT500** Developer Controlled, AI Accelerated Software Development