From comparing the big three AIs, choosing models, understanding the product lineup, to mastering Claude Code — everything you need in one read.
Each of the big three has its strengths — like three different pliers in a hardware store. It all depends on what you need to grip.
Excels at research and deep thinking — ask it a complex question, and it maps the full picture before diving in. When coding, it comprehends the whole project first, then delivers in one shot, catching edge cases you didn't even think of.
The go-to for visual design. Drop a single prompt and it generates beautifully polished web pages. Its image generation is precise and commercially viable — if you're doing design, this is the one.
Never overthinks to the point of over-engineering, and never randomly drops the ball. Consistent, high-quality output is its trademark. Want something you can hand off without worrying? Go with Claude.
Claude comes in three tiers — like consumer, professional, and flagship tools at a hardware store. Just pick the one that fits.
| Haiku | Sonnet | Opus | |
|---|---|---|---|
| Positioning | Lightweight & fast | Daily workhorse | Highest quality |
| Best for | Simple classification, summaries | Everyday task execution | Writing, coding & high-quality work |
| Speed | Fastest | Fast | Slower |
| Cost | Low | Medium | High |
| Free tier quota | Generous | Moderate | Very limited |
| PRO | MAX 5x | MAX 20x | |
|---|---|---|---|
| Monthly (USD) | $20 | $100 | $200 |
| Approx. TWD | ~600 | ~3,000 | ~6,000 |
| Opus quota | Limited | 5× | 20× |
Claude has three entry points: New Chat / New Task / New Session — three entirely different product philosophies. Like a screwdriver, a power drill, and a CNC machine — same brand, vastly different uses.
A classic chat interface — just ask. It can also connect to external tools like Figma and Notion via MCP.
Assign a task and let it run in the background. It can operate your computer — browse the web, organize files, fill out forms, schedule automations — and notify you when done.
A fully autonomous loop — reads files, edits files, runs terminal commands, searches the web, connects services, and works end-to-end without stopping.
| Feature | App | Cowork | Code |
|---|---|---|---|
| Answer questions | ✓ | ✓ | ✓ |
| Read local files | ✕ | ✓ | ✓ |
| Edit local files | ✕ | ✓ | ✓ |
| Operate your computer | ✕ | ✓ | ✓ |
| MCP support | ✓ | ✓ | ✓ |
| Skill support | Usable but not persistent | ✓ | ✓ |
| Background / scheduled runs | ✕ | ✓ | ✓ |
| Learning curve | Low | Medium | High |
Claude is more than just a chat box — its surrounding tools line up like accessories on a hardware store pegboard.
Both products support Tasks, but work differently. Cowork runs locally (your computer must stay on); Code supports cloud scheduling, running even when you're away.
🔧 Think of it as: assigning a to-do, going about your day, and checking the report when it's done
Use plugins to connect Telegram / Discord to a Claude Code session, so you can command it remotely from your phone.
🔧 Think of it as: texting your home computer assistant from your phone to get things done
Persistent cross-device conversations — assign tasks from your phone, execute on Mac, get push notifications on completion. macOS only for now.
🔧 Think of it as: directing your office assistant no matter where you are
More than just reading web pages — it can navigate, click, fill forms, manage multiple tabs, and record workflows.
🔧 Think of it as: an assistant sitting next to you, operating the browser on your behalf
Interactive charts pop up right in the conversation — compound interest curves, periodic tables, concept diagrams, generated on the fly.
🔧 Think of it as: quick whiteboard sketches drawn mid-conversation
Directly operates your desktop — opens apps, controls the browser, fills forms, organizes files.
🔧 Think of it as: a remote assistant who can see your screen and control your mouse and keyboard
Claude Code was originally built for engineers, but because it can read files, edit files, and run terminal commands — the potential is literally limitless. In short: regular Claude is AI that "talks"; Claude Code is AI that "does."
Non-programmers can use it just as powerfully:
Reads your notes folder, auto-categorizes and tags everything
Writes articles based on your note library — your material, your style
Generates slide decks directly from your data
Turns articles into social media posts with captions in one click
Creates visual designs with HTML, integrates with Figma MCP
Downloads YouTube videos, compresses audio, batch-renames files
Follow along — you'll be up and running in ten minutes. Easier than assembling IKEA furniture.
You need Pro or above to use Claude Code. $20/month — great value.
On Mac, press Cmd + Space and type "Terminal" to open it.
curl -fsSL https://claude.ai/install.sh | bash
Create a new Claude-workspace folder and drag it into your editor.
Go to Terminal → New Terminal, then type claude.
Once you see the welcome screen, you're all set! Just talk to it in natural language.
The protocol for AI to communicate with external tools — the USB-C of the AI world. Install the Notion MCP and it can create pages directly.
🔌 Think of it as: a USB-C port — plug in anything
An instruction manual written for Claude. Scripts, examples, and workflows all bundled into a SKILL.md, triggered with a single command.
📋 Think of it as: an SOP manual — follow the steps, no mistakes
Your custom handbook for Claude. It's auto-loaded every session, so Claude always knows your rules.
📖 Think of it as: an employee handbook — read on day one
Claude's own notebook. It remembers your corrections, preferences, and past pitfalls.
📓 Think of it as: a work journal — the more experience, the better the performance
/model — Switch models/compact — Compress old conversation/clear — Clear context/account — Check remaining quota/init — Auto-create CLAUDE.md/resume — Resume last conversation/plugins — Browse & install pluginsType @filename to bring a specific file into the conversation. Highlighted text automatically appears in the input box.
💡 Claude doesn't read all files every time — specify what you need for precision
Each edit prompts an authorization confirmation. You can switch to "auto-edit" mode — four permission levels from strict to fully open.
⚠️ skip-permissions bypasses all safety checks — only recommended in isolated environments
Press ESC to stop immediately. The undo feature lets you: revert only the conversation, revert only file changes, or revert everything — like branching in Git.
💡 Change your mind anytime — zero-cost experimentation
Each conversation has a capacity limit. When usage reaches 70–80%, run /compact to compress old content.
💡 The conversation is like a whiteboard — wipe it when full; CLAUDE.md is the bulletin board on the wall — always there
Shift + Enter for a new line without sending; Enter to send. Organize your thoughts first for higher quality output.
Set a condition, and it fires every single time. Auto-notify on task completion, auto-lint before saving.
🚪 Think of it as: an automatic door — someone approaches, it opens
Create persistent specialist assistants via /agents, with pre-configured tools, models, and memory.
👷 Think of it as: trained specialist employees, ready when you call
Split work across multiple AI agents running in parallel; results are automatically merged.
🏥 Think of it as: turning one exam room into multiple rooms seeing patients simultaneously
Drafts a complete blueprint, lists key questions for confirmation, then executes only after approval.
📐 Think of it as: drawing the blueprint and getting sign-off before breaking ground
Browse all plugins via /plugins. Pre-bundled combos of Skill + Hooks + MCP — all set up in one go.
🧰 Think of it as: buying a toolkit instead of picking items one by one
Describe what you need in natural language, and Claude auto-generates a custom Skill for you.
🍳 Think of it as: nailing a recipe, then writing it down for next time
Start with "what I want" instead of "what you should do." Give it the destination and let it figure out the route.
Don't spoon-feed it step by step — give it the entire workflow at once. That's what it does best.
Try once, save it, use it many times. A good prompt deserves to be packaged.
Collect assets over time — writing style references, design examples, common formats — and deploy them with a single sentence.
It's not just the tool getting smarter — your entire system is evolving. CLAUDE.md, Skills, MCP — they all compound.