Learn Claude Code: The Complete Operator's Guide

What is Claude Code, really?

Open Claude. Ask a question. Get an answer. Close the tab. That is how most people meet it. That is also the floor of what it can do.

Claude Code (the brand) is the same model with the chat box turned inside out. An operating system for AI work, not a tab you visit. To make sense of it, hold two axes in your head:

• Where you work: Web, Desktop, or Terminal. These are the three surfaces Claude ships in.
• How you work: Chat, Cowork, or Code. These are the three modes inside Claude Desktop.

Pick one from each axis. Some combinations exist, some do not.

The three surfaces (where you work)

• Claude Web at claude.ai. Browser-based. Chat only. The version most people meet first.
• Claude Desktop, the standalone app for Mac and Windows. Three tabs inside (Chat, Cowork, Code), connectors, plugins, scheduled tasks. This is where most operators live.
• Claude Code (the terminal CLI), a developer-facing command-line tool. Same engine as the Code tab inside Desktop, just exposed for people who prefer the terminal. Optional if you have Desktop.

Which surface supports which modes

A note on the "Code" naming overlap. Anthropic uses the word "Code" in two related places. The terminal CLI is officially called Claude Code. Inside Claude Desktop there is also a tab called Code. Both run the same engine. The CLI is just the standalone version for people who live in the terminal. If you install Desktop, you already have the Code engine, and the CLI is optional.

You do not have to pick one. Most operators we train at our Claude Code workshop (Auckland and Wellington runs) use all three. We jump between Desktop and the terminal CLI half a dozen times a day at Waboom.

If you have never installed Claude Code on a real machine before, start with our free Claude Code 101 course. Six modules. Two hours. You will leave with a live landing page on the internet. (We built ours over flat whites in Mt Eden. So can you.)

Pick Your Model (and the Plan That Goes With It)

Claude is not one model. It is a family of three, each tuned to a different shape of work, each sitting on a different price/performance curve. Operators who do not know the difference end up paying too much for trivia or paying too little for real engineering.

The three models, in plain English

Which plan you actually need

Anthropic publishes four consumer plans at claude.com/pricing:

• Free ($0): access to Sonnet for chat. No Claude Code. Fine for tasting the product.
• Pro ($17/month annual, $20 monthly): Sonnet for everything, Opus for harder asks, Haiku for cheap. Claude Code included. This is where most operators should start.
• Max (from $100/month): Two tiers, 5x or 20x more usage than Pro. If you hit Pro's limits because you keep Cowork running scheduled tasks or you run Opus a lot, this is the upgrade. We run Max at Waboom.
• Team ($20-25/seat) and Enterprise ($20/seat with usage scaling): for organisations deploying Claude across a team.

The simple rule

Start on Pro at $20 a month. Default your work to Sonnet. Switch to Opus when you hit a hard problem (a complex refactor, a real architecture decision, a Ralph Loop on a big PRD). Switch to Haiku when you are calling Claude programmatically and quality matters less than cost.

We spent six weeks on Pro before upgrading to Max. Hit the limit twice in a week, both times running an Opus session on something we should have planned better. Pay attention to what tripped you up, not just the count. (Me: "It's the model." Also me, two hours later: "It was my prompt.")

If you hit Pro's usage limit twice in a week, upgrade to Max 5x. If you hit Max 5x's limit, you are probably running Cowork autonomously enough that Max 20x or Team pays for itself.

Who Should Use What?

Four common starting points. Find the one that sounds most like you, then follow the "first action" line. You can always graduate to a wider stack later.

If none of these fit exactly, the safe default is the second one: Desktop on Pro at $20/month. It covers about 80% of operators we train at our Claude Code workshop.

The Operator's Stack

Underneath every powerful Claude Code workflow is the same four-layer stack. All four working together is the moment Claude stops feeling like a chat tool and starts feeling like a coworker.

Layer by layer:

Layer 1: Skills (the workflow layer)

A skill is a folder that teaches Claude a workflow once so you stop re-explaining yourself. Write in your voice. Run your team's code review. Generate your weekly report. Anything you have re-explained three times to Claude is a skill candidate.

Skills are the compounding layer of the stack. Most operators we train at the Claude Code workshop get their first working skill running before the end of hour one.

Inside Claude Desktop, click the + next to Skills and hover over Create skill. You get three ways to actually build one:

Create with Claude

The easiest path. A chat opens. You describe the workflow in plain English. Claude asks 2-3 follow-up questions, drafts the skill, saves it. No file management, no dialogs, no terminal. Good for your first skill.

Write skill instructions

A short dialog with three fields. Name (kebab-case, like weekly-report). Description (what the skill does plus when it should trigger). Instructions (the actual prompt-like body). You stay in control of every word.

Upload a skill

For skills you already have. Zip the folder containing the SKILL.md (plus any optional scripts, references, or assets), upload, done. This is how you share skills across teammates or install one you grabbed from GitHub.

All three produce the same thing: a working skill in your account that triggers automatically when a chat matches its description. Pick whichever fits your hour.

Deep dives:

• What is a Claude Skill (plain-English explainer)
• Claude Skills vs MCP: which one do you need?
• Build your first Claude Skill in 15 minutes
• Free Skills 101 course (6 modules, 1 hour)

Layer 2: Connectors (the access layer)

Connectors plug Claude into the tools you already use. Notion, Gmail, Linear, Google Drive, HubSpot, Slack. Once connected, Claude can read your meeting notes, draft replies to your inbox, summarise this week's sprint.

If you know what an MCP server is, Connectors are the user-facing wrapper around that. If you do not, do not worry. They are just toggles in Settings.

Deep dive:

• Connectors 101: plug Notion, Gmail, and Linear in 10 minutes

Layer 3: Projects (the workspace layer)

Anthropic's own description is direct: "Projects in Claude Cowork let you group related tasks into dedicated workspaces with their own files, context, instructions, and memory."

A Project is the folder Claude works inside. It bundles together five things:

• Instructions: your tone, formatting, rules, anything Claude should follow inside this Project
• Files: reference docs you attach so Claude can read them on every task
• Context: a local folder, a linked chat project, or a URL Claude treats as the working source
• Scheduled tasks: recurring Cowork jobs scoped to this Project
• Memory: Claude remembers what happened in earlier tasks within the same Project

When you create one, Cowork offers three starting points:

• Start from scratch: a fresh folder with instructions and files
• Import a project: bring across an existing Claude Chat project into Cowork
• Use an existing folder: point Claude at a folder you already work from on your computer

The setup dialog itself is short, name, instructions, attached files, save location:

One Project per retainer client. One Project per repeating workflow. The brief and reference files become the working memory the next task starts from, so you never re-explain who you are or what you build.

Deep dive:

• Claude Projects: beginner setup guide

Where the Work Happens: Chat, Cowork, and Code

The four capabilities above run inside Claude. But Claude itself has three different ways of working with you, and picking the right one for the task changes everything. Inside Claude Desktop you will see three tabs at the top: Chat, Cowork, and Code.

Anthropic's own visual of Cowork: a Progress checklist running through tasks, with a Context panel pinning the files and skills it has access to. Source: anthropic.com/product/claude-cowork.

Cowork: work delegated, deliverable returned

Anthropic's description of Cowork is direct: "Cowork handles tasks autonomously. Give it a goal and Claude works on your computer, local files, and applications to return a finished deliverable."

This is the mode for delegated work. Hand Claude a goal, a folder of files, maybe a connector or two, walk away. It reads, drafts, organises, synthesises, returns a finished thing. The same capability the developer-facing Claude Code has, simplified for non-technical knowledge work.

Schedule a Cowork task to run weekly and the Monday-morning briefing writes itself. This is where Skills earn their compounding. A scheduled Cowork task running a voice-DNA-enforced skill on your inbox is what an "AI coworker" actually looks like.

Deep dives:

• Claude Cowork: what it is, when to use it (and when not)
• 5 boring jobs to automate with scheduled Cowork tasks

Inside Claude Code: Plan Mode and the 5 Prompts That Change Everything

Most operators install Claude Code, type their first prompt, get a half-built thing back, and conclude AI is not ready. (We did the same thing in week one. Built nothing useful. Blamed the model.)

The real problem: they treated Claude Code like a chat. It is not a chat. It is a junior engineer who needs a brief and a definition of done.

This section is the workflow we drill in the first hour of every Claude Code workshop. Get this right and Claude Code stops feeling unreliable.

Plan Mode: make Claude think first, build second

Claude Code has a mode called Plan Mode. In Plan Mode, Claude reads your brief, asks clarifying questions, proposes a technical plan, lists the files it will create or change, and waits for your approval before touching anything. Nothing builds until you say go.

Turn it on by clicking the dropdown next to the + button at the bottom-left of Claude Code and selecting Plan mode. Or hit Shift + Tab to toggle it.

The trick: in Plan Mode, drop your raw brainstorm. Bullet points, fragments, half-thoughts. Claude organises it into a clean brief using the Role / Task / Format / Context framework and proposes the architecture. You did not need to know what Next.js or Tailwind is. Claude chose the right tools, you approve, then Claude starts building.

The 5 prompt checks (the 30 seconds before you hit Enter)

Before every non-trivial prompt to Claude Code, run through these 5 checks. They take 30 seconds and they 10x the quality of what comes back.

Memorise Scope, Diagnose, Explore, Define, Verify. Whisper it before every prompt. Inside three weeks it becomes muscle memory and the output quality jumps without you trying harder.

Bonus: make CLAUDE.md do the heavy lifting

Once you have the 5 checks in muscle memory, push the recurring parts (scope rules, voice, "never do this", definition of done) into a CLAUDE.md file at the root of the project. Claude reads it on every prompt automatically. The 30 seconds before Enter shrinks to 5 seconds.

Choose Your Mode

Inside Claude there are three buttons most people walk past every day: Extended Thinking, Web Search, and Research. Three different jobs.

• Extended Thinking for deliberation. Hard reasoning, multi-constraint problems, design decisions.
• Web Search for live facts. Anything that changed since training cutoff.
• Research for reports. Multi-source synthesis on a topic, with citations.

Thinking is for deliberation. Search is for facts. Research is for reports. That is the whole map.

Power-ups for Operators

Once the stack is wired up, these are the operator-level patterns that separate "has a Claude account" from "runs a real AI workflow".

The CLAUDE.md file

Drop a CLAUDE.md file at the root of any Claude Code project and Claude reads it on every prompt. Project rules. Conventions. Style. The "never do this" list. Anything Claude needs to know to work in that folder, captured in one file.

Two hundred lines of CLAUDE.md replaces an hour of re-explaining yourself every session. Drop the file at the root of your project, list your stack, your style rules, your "never do this" constraints, and Claude reads the lot on every prompt.

We write CLAUDE.md files live with operators in our Claude Code workshop. The first one usually takes 20 minutes. Every Claude Code project you ever start after that gets one.

The Ralph Wiggum Loop

Waboom's name for an autonomous agent loop that grinds a complex task to completion. Give it a product requirements doc, walk away, come back to a finished feature. It is overkill for small tasks and devastatingly effective on big ones.

We demo Ralph live in the Claude Code workshop. Operators usually watch one Ralph run and start spotting candidates everywhere.

Install Ralph yourself:

Ralph is a custom Claude Code connector. Open Claude Code, then go into the Customize panel:

1. Click the + next to Personal plugins, then Add marketplace.
2. Paste this GitHub URL: https://github.com/snarktank/ralph and click Sync.
3. Find “Ralph skills” by snarktank in the plugin list and click Install.

Trigger a Ralph run with /ralph-skills:ralph followed by the goal you want it to grind to completion, then ralph.sh commence to kick it off.

In Your Tools

Claude is not just a chat window. It plugs into the apps you already live in. Excel, Word, PowerPoint, Outlook all have official Claude add-ins now.

The biggest unlock for non-developers is Claude for Excel. Anthropic describes it as "an add-in that integrates Claude into your Excel workflow, designed for professionals who work extensively with spreadsheets, particularly in financial analysis and modeling."

Point Claude at a workbook and ask in plain English: build a pivot, clean column C, write the formula, explain what this sheet does. Every change goes through a permission dialog so you stay in control.

What Claude for Excel can actually do

• Read and explain: ask questions about your workbook and get answers with cell-level citations
• Update assumptions: change values and watch Claude maintain every formula dependency
• Build and fill templates: create models from scratch or fill an existing template from a 10-K or invoice PDF
• Debug errors: find every #REF, #VALUE, and circular reference in a workbook and explain why
• Track every change: Claude highlights every cell it touches with explanatory comments
• Edit and format natively: sort, filter, edit pivots and charts, apply conditional formatting, set data validation
• Use your Skills and Connectors: any skill you have enabled in Claude works inside Excel too

The permission flow that keeps you safe

Claude warns before overwriting existing data. Dangerous Excel functions (WEBSERVICE, IMPORTDATA, INDIRECT, CALL, EVALUATE, file-system functions) require explicit approval via a confirmation pop-up before they execute.

For the first month keep "allow this action" as your default rather than "allow all edits". Graduate slowly. Spreadsheets break easily and the dialog is what stops a misphrased prompt from corrupting your model.

Anthropic's explicit guidance: "Only use Claude for Excel with trusted spreadsheets, not spreadsheets from external untrusted sources."

A few real prompts that work

• "Build a three-statement model for a SaaS company."
• "Compare actuals to budget and explain the largest variances."
• "Standardise phone numbers in column F to +1 (XXX) XXX-XXXX format."
• "Find all #REF errors in this workbook and trace each one back to the source cell."
• "Extract the financial table from this PDF into a new sheet."

What it does not do

• No support for VBA, macros, or data tables (the legacy ones)
• Only .xlsx and .xlsm files
• Available on Pro, Max, Team, and Enterprise plans only (not free tier)
• Excel on the web, Windows, Mac, and iPad. Not Android. Older Excel 2016/2019 perpetual licences not supported.

Anthropic explicitly cautions: do not use Claude for Excel for final client deliverables without human review, audit-critical calculations without verification, or models containing highly sensitive or regulated data without proper controls.

How We Use This at Waboom AI

Everything above is theory until it ships work. Three workflows we run at Waboom every week:

This is the exact playbook we teach in the room at our Claude Code workshop. Three hours, bring real work, leave with the same operating stack running.

Two Ways to Learn It

Whichever fits your style.

Self-paced: our two free courses cover the foundation. Claude Code 101 walks you from zero to a deployed landing page in six modules. Claude Skills 101 teaches you to build, debug, and share your first skill. Together they are about three hours.

Hands-on: the Claude Code workshop is three hours of live training. You bring a real problem, leave with it shipping. NZ$299 + GST, money-back guarantee, dates monthly across New Zealand and Australia.

Most operators do the free courses first, then come to the workshop with sharper questions. Either way, by the end of next week you stop re-explaining yourself to Claude, stop copy-pasting between tools, and start shipping work you would have charged an agency for.

Common questions we get asked

Pulled straight from operator questions at our last six Auckland and Sydney workshops.