06
Setup AI Fundamentals

Claude Code: Setup, Modes & Memory

A terminal-native agent loop with shell access. Five surfaces decide its behaviour.

One Agent Loop, Five Surfaces That Steer It

The host runtime owns your folder. Permission modes gate every action. Memory survives between sessions. Skills, sub-agents, and MCPs add capability.

Host Runtime Claude Code CLI Read → Decide → Act
Haiku 4.5 Sonnet 4.6 Opus 4.7
1
Skills & Slash Commands

Reusable workflows triggered in one keystroke.

/plan /init /compact
2
Sub-Agents

Specialised Claude instances Claude can delegate to.

Research Review QA
3
Memory Layers

Four scopes concatenate at session start.

Managed User Project Local
./CLAUDE.md
4
MCP Tools

External plug-ins: databases, browsers, APIs.

Postgres Playwright GitHub
5
Permission Modes

Six modes. Cycle with Shift+Tab. Detail below.

6 modes Shift+Tab

Pick The Mode That Matches The Trust

Plan first, approve, then accept edits. Make Shift+Tab into plan mode muscle memory. dontAsk and auto exist; reach for them only after the basics are second nature.

Mode 01

default

Reads only. Everything else prompts you, action by action.

Day One · Sensitive
Mode 02

plan

Reads, writes a plan, no source edits. Approve the shape before any action runs.

Default Habit
Mode 03

acceptEdits

Reads + file edits + basic FS auto-run. Shell commands still prompt.

Trusted Direction
Mode 04

bypassPermissions

Everything. No checks. Refuses to start as root. Sandboxed containers only.

Throwaway VM
100tasksai.com · © 2026 100 Tasks AI