Reference Guide

When to Use What: Claude, Claude Code, and Fable

A working reference for choosing between Claude's chat interface, Claude Code (the terminal tool), and the different model tiers — written after using all of them on the Market Prediction Engine build.

Chat vs. Claude Code — Different Tools, Different Jobs

SurfaceWhat it's forKey limitation
Claude.ai chat (this) Conversation, planning, web research, editing files via the GitHub API one call at a time, generating documents. Runs in a sandboxed environment with a restricted network allowlist — can't reach every domain (e.g. GitHub's own Actions log storage lives outside it).
Claude Code Lives in your terminal (or a desktop app, if you'd rather skip the terminal entirely), reads your actual local repo, edits real files, runs real commands, has normal internet access. Requires a Claude Pro/Max subscription or API credits — not available on the free plan.
Rule of thumb: use this chat for planning, research, and one-off file edits via API. Use Claude Code when you want it working directly inside your project folder, running commands, and iterating across many files in one sitting.

Installing Claude Code — Step by Step

Two options: terminal (most flexible) or the desktop app (no terminal at all).

Option A — Desktop app (easiest, no terminal):

  1. Download the Claude Code desktop app for your OS from claude.com
  2. Install and open it, sign in with your Claude account
  3. Open your project folder from inside the app

Option B — Terminal (macOS example):

  1. Open Terminal: press Cmd + Space, type "Terminal", press Enter
  2. Paste this command and press Enter: 
    curl -fsSL https://claude.ai/install.sh | sh
  3. Wait for "Claude Code successfully installed!"
  4. Type claude and press Enter — a browser window opens for login
  5. Sign in with the same account you use for claude.ai

Windows: Open PowerShell, and if it says "irm is not recognized," you're in the wrong shell (that's a CMD vs. PowerShell mixup — the install page gives you the exact right line for each).

Requires a Claude Pro ($17–20/month) or Max subscription, or an API account with credits — the free claude.ai plan doesn't include Claude Code. Check current pricing at claude.com before committing to a tier.

Your First Session

  1. Navigate to your project folder in Terminal (or open it directly in the desktop app), e.g. cd market-prediction-engine
  2. Run claude — it reads any CLAUDE.md file in that folder automatically, so it already knows the project context
  3. Start small: ask it to "explain what this repo does" or "find the entry point" before asking for changes — watch it read files and reason before trusting it with edits
  4. Press Shift+Tab to enter Plan Mode — it proposes a plan without touching any files, you approve, then it executes
  5. Type /help anytime to see all available commands
  6. Type exit or press Ctrl+D to leave
You don't need to know how to code to use it — you're directing it in plain English, the same way you'd direct me in this chat.

Which Model, When

ModelBest forCost tier
Haiku 4.5 Lightweight, high-frequency, repetitive tasks — formatting, simple lookups, grunt work inside a larger workflow cheapest
Sonnet 5 Your default for almost everything — routine development, most of a normal build, day-to-day chat low
Opus 4.8 Complex architecture decisions, tricky debugging, system design — a step up when Sonnet is genuinely stuck moderate
Fable 5 Frontier-tier — large ambiguous builds, long-running multi-file work, a bounded end-to-end review pass. Reserve for one specific task, not everyday use. highest
Switching models mid-session makes the new model re-read the entire conversation history — more expensive than starting fresh. If you're deliberately stepping up to Fable for one big task, start a new session for it.

Cost Control Habits

  • Set Sonnet as your default in ~/.claude/settings.json: { "model": "sonnet" } — so you never accidentally run the expensive model for routine work.
  • Use Plan Mode (Shift+Tab) before big changes — see the scope and cost before committing, not after.
  • Break large asks into steps. "Build a full auth system" is expensive and error-prone in one shot; "create the schema," then "add the endpoint," then "build the form" is cheaper and easier to review.
  • Write a CLAUDE.md in each project so Claude Code has context on day one instead of you re-explaining it every session (token cost, not just your time).
  • Run /clear between unrelated tasks in the same session to keep the context lean.

Applied to This Project

The actual plan for the Market Prediction Engine's frontend MVP:

  1. Install Claude Code, default to Sonnet 5
  2. Build the MVP frontend entirely on Sonnet — a dashboard reading data/predictions/, data/trust/, and data/logs/ directly from the repo
  3. Use Plan Mode before any multi-file step
  4. Once the Sonnet build works, open a fresh session on Fable 5 for exactly one bounded task: a full end-to-end review with test-writing

See the Build Log for live status, and the Team 1 Dissection for what the original student project actually did, sourced directly from its repo.

Listen to This Page
Ready.