Claude Code

The cheat sheet - from your build session, right next to your keyboard
October
2026

Keep this open in a tab or print it · saves you searching those first weeks

1 Getting started

  1. Open your terminal in your project folder.
  2. Type claude and hit enter.
  3. Run /init once - Claude reads your project and creates an instruction file. Name it AGENTS.md and other AI tools read it too.
  4. Run /statusline once - from then on you always see how full your context is.
  5. Then just ask, in plain English. Build, explain, fix - anything goes.
AGENTS.md = your project's memory. Put agreements, style and ground rules in here; Claude reads it automatically every session.

3 Keys to know by heart

Shift+TabSwitch permission mode (see block 5)
EscInterrupt Claude mid-answer
Esc EscOpen the rewind menu
@Reference a file or folder
↑ ↓Browse your previous prompts
Ctrl+RSearch back through your prompt history
Ctrl+VPaste an image or screenshot (Mac: Ctrl, not Cmd. Windows: Alt+V)
Ctrl+C Ctrl+CTwice = quit Claude Code
?Show all shortcuts

4 MCP & CLI - your new colleagues

MCP
An expert colleague you install - knows the documentation, operates a service's API. E.g. the Cloudflare MCP (block 8).
CLI
Command line interface - tools driven by text commands; Claude talks to these at lightning speed. E.g. the GitHub CLI (gh).

2 The commands you'll actually use

/clear
Start fresh. New topic = clear first.
/compact
Shorten a long conversation without losing the thread.
/usage
What you have used and how much of your plan is left.
/rewind
Go back to an earlier point. Wrong turn? Rewind.
/resume
Pick up an earlier conversation.

5 The 4 modes (Shift+Tab)

Manual - politely asks permission for every action. Safe starting point.
Accept Edits - approves changes automatically. Use once you trust it.
Plan - read-only. Thinks along, touches nothing. Ideal before a big step.
Auto - keeps going without asking each time. The starting mode in recent versions.

6 How to get the most out of it

  1. Give context, not just an instruction: what, what for, for whom.
  2. For big things, switch to plan mode first (Shift+Tab) - read the plan, then approve.
  3. Work in small steps. Finished a piece? /clear and move on.
  4. Unsure about a change? Esc Esc and rewind. Nothing breaks.
  5. Slop alarm: ask for character, give examples. "Don't make it generic."
Beyond the basics Cheat sheet · page 2

7 Skill or agent?

SkillAgent
What Knowledge + instructions in a file (SKILL.md): what AI knows. An acting entity that executes, decides, iterates: what AI does.
Nature Passive - loaded when relevant. Active - a chain of steps: input → logic → output, with tool access.
E.g. Your house style, review checklist, quoting conventions. A helper that works out a subtask while you carry on.
Rule of thumb: knowledge you want to reuse → skill. A job that must run on its own → agent. They combine well: an agent loads skills.

8 The stack to build with

Nuxt
Framework - site, app and API in one project. Claude knows it inside out.
Tailwind
Styling - design as utility classes. AI writes it fluently and consistently.
Cloudflare
Hosting - Pages & Workers, free to start. Hook up the Cloudflare MCP and you're live in seconds.
GitHub
Version control - every step recorded; Claude drives it via the GitHub CLI (gh). Home base for your skills (block 9).
Tip: create a design.md with your brand style and assets, and share it like your skills (block 9) - everything you build is instantly on-brand.

9 Sharing skills across your organisation

  1. Put your skills in one git repo: skills/<name>/SKILL.md.
  2. Everyone clones it, e.g. to ~/git/skills.
  3. Link with a symlink from the git folder: ln -s ~/git/skills/<name> ~/.claude/skills/ - the skill works in every project and the repo stays the source.
  4. Colleague built something new? git pull and you have it too.

10 Superpowers

/skills
Skills = ready-made recipes. One command, whole job done (e.g. a GEO/SEO check on your site).
/skill-creator
Claude helps you make a skill. Install once: /plugin install skill-creator@claude-plugins-official
/agents
Subagents - ask Claude for a helper that works out a subtask while you carry on.
/mcp
Connect external sources (Stripe, GitHub, your analytics) so Claude can reach them directly.
/memory
Edit what Claude remembers about your project.
/model
Switch models (e.g. a heavier model for hard jobs).
/voice
Talk instead of type - sharing thoughts out loud beats typing them. Great for brainstorms and giving context.

11 Which model when?

Opus 5.5
The default - strong on everyday and complex work. Start here.
Sonnet 5.5
The workhorse - routine tasks, fast and still very capable. Switch: /model.

12 Thinking level (effort)

/effort
Dial how thoroughly Claude thinks: low → medium → high → xhigh → max. Simple job? Lower saves time and money. Default: medium.
ultrathink
Type it right in your prompt: one-off extra-deep thinking for that answer, without changing the session.

13 Build a second brain

  1. Transcribe everything - meetings, calls, loose thoughts. Dictation: Paraspeech (Mac), Oorapp (free, beta) or Handy (Windows, free). As a team: Wispr Flow or Fireflies (paid, cloud).
  2. Store the transcripts as text files in a folder in your project.
  3. Feed them to your agents - every meeting becomes context to build with, every thought searchable.
Recording → transcript → context → build. Your conversations become fuel instead of memories.