1. In the Claude app, with a Claude subscription
2. In your terminal (the CLI): sign in with a subscription, or use an API key
3. On the workshop website: add your API key to your Profile

Tonight we use the workshop website with your own API key.

• In a chat, you copy and paste code back and forth
• Claude Code works in your files and your terminal, and does the work itself
• It is an agent: software that takes actions to reach a goal you set
• It can read your code, edit files, run commands and search the web

1. PromptYou ask in plain English
2. Gather contextIt reads files and searches
3. ActIt edits files and runs commands
4. VerifyIt checks the result, then repeats

• You can interrupt at any point: press Esc and tell it what to do instead

• Press Shift+Tab to move between modes
• It can make mistakes. Stay in the loop so you catch them early

1. ExploreClaude gathers the context it needs for your project
2. PlanYou agree a plan of action to measure success against
3. CodeBack and forth between you and Claude until it is right
4. CommitReview the change and save it, then start the next one

• Press Shift+Tab until the status bar shows plan mode on
• Claude reads files and runs commands to explore, but edits nothing
• It asks questions, then comes back with a plan you can change
• This is the cheapest place to change course: no code is written yet
• Say what “correct” looks like, so Claude knows when it is done

• /diff shows the change line by line: what was removed and what was added
• Look for changes you didn’t ask for, tests that got weaker, new packages and hard-coded values
• All wrong? /rewind (or Esc twice) takes you back
• /code-review is a second opinion from a clean context, with none of your session’s history

• Context is Claude’s working memory: your messages, its replies, every file it read, every command’s output, plus CLAUDE.md
• Think of it as a desk with a fixed size
• Claude can only work with what’s on the desk
• When it fills up, older details fall off and answers get vaguer

• Claude starts every session fresh. It reads CLAUDE.md at the start of each one
• In your project folder it is shared with the team. ~/.claude/CLAUDE.md is just for you, in every project
• Put in commands, conventions, “don’t touch this folder”
• Correcting Claude over and over? Ask it to save the rule
• /init asks Claude to write or improve the file. Read it before you accept

• /context shows what’s using space
• /compact swaps the conversation for a short summary. Use it mid-task when you need to keep going
• /clear empties the conversation. Use it when you start something new
• Be specific: a vague prompt makes Claude explore more, and that costs more space
• Switch off MCP servers you aren’t using, and hand legwork to a subagent

• Exploring your code and searching the web fills up context fast
• A subagent is a helper Claude starts for one job, like “explore this codebase for me”
• It has its own context window and does the legwork there
• It returns only a summary, so you get the answer without the journey

• A subagent is a Markdown file with a short settings header (YAML frontmatter)
• .claude/agents/ works in this project
• ~/.claude/agents/ works in all your projects
• Easiest way: describe the subagent you want and ask Claude to write it
• Try it: ask Claude to use a subagent to explore your project

• A folder of instructions and examples that teaches Claude how to do one thing
• Only its name and description sit in context until Claude uses it
• CLAUDE.md is for facts Claude always needs
• A skill is for a procedure it needs sometimes
• A prompt is for a one-off request

Use the elite-web skill to add an About page at /about that says what this site is for,
and link to it from the home page.

• Claude matches your request against each skill’s description and loads the ones that fit
• Naming the skill, or typing /elite-web or /elite-react, removes the guesswork
• Your starter ships its skill in .claude/skills/

.claude/skills/site-check/
└── SKILL.md

• SKILL.md starts with a header that gives a name and a description
• In the description, say what the skill does and when to use it
• .claude/skills/ is part of the project: anyone who clones it gets the skill
• ~/.claude/skills/ is personal and works in every project

• Out of the box, Claude can read files, run commands and fetch web pages
• MCP (Model Context Protocol) servers give it more tools: a database, an issue tracker, a cloud account
• HTTP servers are remote services, hosted by the provider
• Stdio servers are local programs that run on your machine

claude mcp add --transport http --scope project cloudflare-docs https://docs.mcp.cloudflare.com/mcp

• Start claude again, approve cloudflare-docs, then type /mcp. It should say connected

claude mcp add --transport http --scope project cloudflare https://mcp.cloudflare.com/mcp \
  --header 'Authorization: Bearer ${CLOUDFLARE_API_TOKEN}'

• Keep the single quotes, so the real token never ends up in .mcp.json
• Every connected server takes some context. /mcp shows what’s connected and lets you disable it
• A command-line tool such as gh can do the same job for less context

• CLAUDE.md and skills are advice. Claude follows them almost always
• A hook is a command Claude Code itself runs at a fixed moment, every time
• Claude can’t skip it or talk its way past it
• Use one to format files after edits, block dangerous operations, or notify you when Claude finishes

In settings.json you pick an event, an optional matcher (which tools it applies to) and a command to run.

• /hooks lists the hooks you have configured

• The hook gets the tool’s name and input, such as the command or the file path
• Exit code 0: the call goes ahead
• Exit code 2: the call is blocked, and the message goes back to Claude as the reason
• .claude/settings.json is part of the project, so the whole team gets the same hooks
• A hook runs real commands. Only add ones you have read and understood