# Raycast AI Quick Start

A keyboard-first AI interface with full context, model switching, and local memory.

---

## Why Switch from ChatGPT

ChatGPT's web interface restricts what the model actually sees. Run the [LLM Accuracy Test](https://guide.zenaitutoring.com/llm-test/) to verify - typical results show ~20% document coverage. The interface lies about having "full access."

Raycast AI:
- Full context (no silent truncation)
- Paste large text directly from clipboard
- Switch models mid-conversation
- Local memory you own
- Keyboard-driven (no mouse required)

---

## Installation

1. Download from [raycast.com](https://raycast.com/?via=blake-sims) (Mac/Windows) — 10% off with this link
2. Subscribe to Pro for model access (~same price as ChatGPT Plus)
3. Set a global hotkey: Search "AI Chat" → Cmd+K → Configure Command → Set Hotkey

> **Windows users:** As of February 2026, Raycast AI Chat is not yet available on Windows — but it's coming soon. The rest of Raycast (launcher, snippets, clipboard history) works on Windows today. For now, use [Claude.ai](https://claude.ai) or Claude Code for AI chat on Windows.

---

## Essential Shortcuts

| Action | Shortcut |
|--------|----------|
| Open actions menu | `Cmd + K` |
| New chat | `Cmd + N` |
| Change model | `Cmd + K` → "Change model" |
| Branch chat | `Cmd + K` → "Branch chat" |
| Regenerate with different model | `Cmd + K` → "Regenerate with model" |
| Copy response | `Cmd + K` → "Copy response" |
| Copy deep link | `Cmd + K` → "Copy chat deep link" |
| Toggle sidebar | `Cmd + K` → "Toggle sidebar" |

**Remember:** `Cmd + K` is the nervous system. Everything lives there.

---

## Model Switching Strategy

Keep context constant, change only the model:

```
f(context, prompt, model_A) → output_A
f(context, prompt, model_B) → output_B
```

1. Get deep into a problem with one model
2. Branch the chat at a decision point
3. Regenerate with a different model
4. Compare outputs or feed one model's answer to another

Use this to catch hallucinations and develop intuition for model strengths.

---

## Custom AI Presets

Create reusable agents with custom system prompts:

1. Search "AI Chat Presets"
2. Create new preset
3. Set: icon, name, model, system instructions
4. Enable tools (Browser, Web Search, MCP extensions)

Unlike ChatGPT, there's no character limit on system instructions.

---

## Local Memory (MCP Server)

### Prerequisites
- Node.js installed (for `npx`)

### Install

1. Copy this config:

```json
{
  "mcpServers": {
    "memory": {
      "command": "npx",
      "args": ["-y", "@modelcontextprotocol/server-memory"],
      "env": {
        "MEMORY_FILE_PATH": "/path/to/your/memory.jsonl"
      }
    }
  }
}
```

2. Get your absolute path:
```bash
cd ~/Documents  # or wherever you want to store it
pwd
# Copy output, append: /memory.jsonl
```

3. In Raycast: Search "Install MCP Server" → paste config → update path

4. Test: New chat → type `@` → select memory → ask it to remember something

### Migrate from ChatGPT

1. ChatGPT → Settings → Personalization → Memory → Manage
2. Copy all memories (no export button, just copy text)
3. Paste into Raycast AI chat with memory tool active
4. Ask it to store the information

Your memories now live locally in a JSONL file you control.

---

## Recommended MCP Tools

| Tool | Purpose | Setup |
|------|---------|-------|
| Memory | Persistent local memory | See above |
| [Exa.ai](https://exa.ai) | Better web search | Free, no API key needed |
| [Context7](https://context7.ai) | Up-to-date code documentation | Requires API key |

Access tools in chat with `@` symbol.

---

## Deep Links

Save chat contexts as URLs:

```bash
# Copy deep link
Cmd + K → "Copy chat deep link"

# Open from terminal
open "raycast://extensions/raycast/raycast-ai/ai-chat?context=..."
```

Store these in notes or scripts to jump back to specific conversations.

---

## Quick Workflows

### YouTube → Summary
1. Open video, press `Cmd + Option + I` (dev console)
2. Paste transcript extraction script
3. Copy transcript → Raycast AI → paste → ask for summary

### Meeting → PRD
1. Copy meeting transcript
2. Open PRD Creator preset (or use [PRD Protocol](https://guide.zenaitutoring.com/prd-protocol.md))
3. Paste → walk through discovery process
4. `Cmd + K` → Copy response → save to project

### Verify Claims
1. Model makes architectural claim
2. Branch chat
3. Regenerate with different model + Context7 tool
4. Compare outputs, feed discrepancies back

---

## Creativity & Reasoning Settings

Access via `Cmd + K`:

| Setting | Low | Medium | High |
|---------|-----|--------|------|
| Creativity | Coding, logic | General use | Brainstorming |
| Reasoning | Quick answers | Default | Complex problems |

Reasoning effort only applies to reasoning models (Opus 4.5, o1, etc).