Contents

What's Inside

Chapter One · Facts checked June 2026

The New Claude — Models, the Heuristic & the Product Map

Let's begin by retiring a question. In 2026 the useful question is no longer "which is better, Claude or ChatGPT?" — it's "which Claude, doing which job, inside which product?" Claude is no longer a chatbot you ask things. It's a tiered family of models wrapped in a suite of products that do the work. Everything in this book hangs off that reframe, so we set it firmly here.

How Claude actually works

Underneath the polish, Claude is a prediction engine trained on a vast amount of text — not a database and not a search engine. It generates the most plausible next words given everything in front of it. That single idea explains its strengths (drafting, summarising, reformatting, reasoning over what you give it) and its one notorious weakness: hallucination. When asked to recall something it was never given, it doesn't lie — it confidently fills the gap. The fix isn't to distrust it; it's to give it the source material and to verify the claims that matter. In 2026 Claude also uses tools — web search, code, your files, your apps — which removes much of the old weakness, but the underlying behaviour is unchanged: good input, a clear job, and a check on what counts.

Four tiers, one heuristic

The model family is a ladder, and each rung trades cost and speed for capability: Haiku (fast and cheap, for high-volume, low-latency work), Sonnet (the balanced workhorse, the right default for most business tasks), Opus (the flagship, for hard reasoning, long documents and multi-step agentic work), and the new Mythos-class tier (frontier capability — and, as we'll see, currently suspended). The whole ladder collapses into one heuristic worth memorising: start on Sonnet, move up to Opus when the task is hard or the stakes are high, drop to Haiku when you just need speed and volume. Model choice is gears; you don't drive everywhere in fifth. The "Which Claude?" picker drills it against real scenarios.

A 72-hour cautionary tale

The Mythos tier teaches a second lesson, free of charge. Anthropic launched the Mythos-class Fable 5 on 9 June 2026 — state-of-the-art on nearly every benchmark — and suspended it for all customers three days later, on 12 June, after a US government export-control directive. Opus 4.8 and everything else stayed up. The business takeaway is concrete and memorable: frontier-model access can be revoked overnight by regulators, so never build a critical workflow on a single model. Design with fallbacks. Capability is thrilling; resilience is strategy.

Advisor to executor: the product map

Now the map you'll walk for the rest of the book. Claude has moved from advisor (it answers your question) to executor (it does the task). On the advisor side sits chat — think, draft, analyse, make Artifacts. On the executor side sit Cowork (the desktop coworker), Claude in Chrome (the browser agent), the Office add-ins (Claude inside Excel, PowerPoint and Word), scheduled tasks, and Claude Design. Binding it all together is the connective tissue — connectors and MCP, Skills, and memory. The interactive product map lets you click each piece. Hold this picture in your head: we walk it clockwise, from "help me write this" to "go do this for me."

Chapter Two

Prompts That Hold Up — the RCTF Framework

Before you can delegate work to Claude, you need output you can rely on. Here is the most useful thing in the whole book for the person who "tried AI once and it was meh": most disappointing output isn't the model's fault — it's an underspecified prompt. "Make it good" isn't an instruction; it's a wish.

The anatomy of a prompt that works

The house framework is RCTF, and it is four boxes. Role — tell Claude who to be ("you're a sharp executive editor"), which sets the vocabulary and the standard. Context — the situation, the audience, the source material; this is where roughly 80% of quality comes from, and it's the box most people skip. Task — one specific, clear instruction. Format — exactly what you want back: length, structure, tone, "five bullets," "a table," "an email under 150 words." The RCTF prompt builder assembles a complete, copyable prompt as you fill the boxes, and shows the difference between a wish and an instruction in real time.

The two habits that compound

Two moves lift quality out of proportion to their effort. First, give an example of what "good" looks like — paste the tone or format you want rather than describing it in adjectives; showing beats telling. Second, name the audience — "write this for a sceptical CFO" silently changes vocabulary, length and emphasis. And a third, slightly magical one: ask Claude to interview you. "Before you write it, ask me three questions that would make this better" turns a one-shot into a quick conversation, and it knows what it's missing better than you do.

Steering beats restarting

Treat the first answer as a draft to react to, not a verdict. "Shorter." "You missed the budget point." "Give me three versions." Claude holds the whole thread, so two or three good follow-ups usually beat a perfect opening prompt — and beat starting over in frustration. The conversation is the tool, not the first reply.

Chapter Three

Making Claude Yours — Projects, Memory, Styles & Artifacts

Two complaints come up every time: "it's useful, but it sounds like AI," and "I have to re-explain everything from scratch each time." Both are solved problems. This chapter makes Claude feel bespoke to you, and turns its answers into things you can actually hand on.

Projects, Memory and Styles

Three features stack to make Claude yours. A Project is a dedicated workspace with its own knowledge: load the relevant documents and a line of custom instructions once, and every chat inside already knows them — you stop re-attaching and re-explaining. Memory gives Claude continuity: it can search and reference past chats on demand and synthesise insights across your history, which you can view, edit, pause or reset. Styles fix the "sounds like AI" problem: feed Claude two or three samples of your real writing and it matches your voice. The shorthand worth keeping: Project = your knowledge, Memory = your continuity, Style = your voice. Together they turn a stranger you brief every time into a colleague who already knows the file. The feature explorer walks all three with setup steps.

Artifacts: from answer to deliverable

An Artifact is a dedicated panel beside the chat where Claude builds something substantial you can see, edit, reuse and share — a formatted one-pager, a comparison table, a working calculator, a chart, a simple web page. It's live and iterative: "add a column," "make it blue," and it updates in place. You're co-editing, not re-generating. Live Artifacts go further, re-running their own code to refresh with new data. The Artifact demo lets you feel what co-editing-by-asking is like.

File creation: the actual file

Distinct from Artifacts, and a big one for business, is file creation: Claude runs code in a private sandbox to produce downloadable .xlsx, .pptx, .docx and .pdf files — available to all users, including the Free tier. "Turn this CSV into a financial model, a one-page memo and a six-slide deck" is a single request. The distinction to keep clear: an Artifact is something you build and iterate inside Claude; file creation hands you a finished file to download. Either way, you leave with the thing, not just advice about it.

Chapter Four · Facts checked June 2026

The Agentic Suite — Cowork, Chrome, Office & Schedules

This is the heart of the book and the place the real return on investment lives. Everything so far made Claude a better advisor. Now it becomes an executor: you hand over whole multi-step jobs and supervise instead of doing.

Cowork: your desktop coworker

Cowork is the clearest example of AI that does the work, not AI that answers questions. It's the third mode in the desktop app: you give it a goal and access to specific folders, and it reads, edits and creates files and completes multi-step work, returning finished deliverables — a formatted spreadsheet, a memo, a briefing — rather than step-by-step advice. The crucial design point is human oversight by design: Cowork shows you a plan and waits for approval before acting, and loops you in before anything significant. That approval gate is your safety belt; never switch it off for work that touches real files or systems. The Cowork walkthrough lets you step through a real session, approval gates and all. It's included on all paid plans (Pro upward), and consumes limits faster than chat.

Acting where you already work

Two more executor surfaces meet you where the work lives. Claude in Chrome is a browser agent that sees your active tab and can navigate, click and fill forms — but it carries the book's sharpest safety lesson: prompt injection, where hidden instructions on a page or in an email try to hijack the agent. Mitigations help (attack success fell from roughly a quarter to around a tenth, and lower on newer configs), but it isn't zero. Keep the browser agent away from banking, passwords and sensitive data, and close sensitive tabs. The Office add-ins put Claude inside Excel, PowerPoint and Word (Outlook in beta): ask about any cell with citations, edit values while preserving formulas, restructure a deck inside your template — with every change highlighted for review, and nothing sent until you click send. The Excel add-in demo shows the review-and-approve loop. This is the direct challenge to Microsoft Copilot — Claude where the hours actually go.

Scheduled tasks: proactive Claude

The last shift is from reactive to proactive. Scheduled tasks run hourly, daily or weekly — a morning brief, a weekly report, recurring research — so the work happens without you starting it. Be honest about the limit: Cowork's scheduled tasks run only while the computer is awake and the app is open. And keep the safe-use rule: start every recurring task as a draft you approve, watch the first few runs, and reserve fully autonomous runs for low-stakes, reversible work. You earn autonomy run by run, exactly as you would with a new hire. The schedule planner helps you compose one safely.

Chapter Five

Wiring Claude In — Connectors, Skills & Design

Module 3 showed Claude doing work. This chapter is the connective tissue that turns a handful of features into a single system — and adds the visual maker most people don't know exists.

Connectors and MCP

Connectors are one-click integrations that let Claude read from and act in the tools you already use — Gmail, Drive, Calendar, Slack, Asana, Jira, Notion, HubSpot, Microsoft 365 and dozens more — so it can pull a document, summarise a thread, check your calendar or draft a reply from your real context. Underneath them is MCP, the Model Context Protocol, an open standard best described as "a USB-C port for AI": it's why Claude connects to so many systems and why the list keeps growing. You don't implement it; you benefit from it, the way you don't need to understand TCP/IP to send email. The governance point matters, though: connecting grants access, and each connector has read-only versus read-and-write permissions you should review at authorisation — use a personal or sandbox account for experiments, and let Team/Enterprise admins govern connectors for real work data. The connector map shows each one's reach and the permission you'd be granting.

Skills: own your best workflow

A prompt is something you write; a Skill is something you own. It's a packaged, reusable capability — instructions, examples and reference files bundled so Claude does a specific job your way, consistently, without re-prompting. You can install ready-made Skills from the marketplace or build your own with the Skill Creator, no coding required. (This very course was built with a custom minisite Skill.) Where it fits the system is the whole point: the recurring valuable work becomes a Skill, the knowledge lives in a Project, the reach comes from Connectors, and the execution happens in Cowork. That's a personal operating system, not a chat window. The Skill Scoper helps you scope your first one.

Claude Design

Finally, the maker most people miss. Claude Design is a visual workspace — chat on the left, a live canvas on the right — that turns a description into real, editable design: a deck, a one-pager, a landing page, a prototype. Describe it, refine it by chat and direct edits, and export to PDF, PowerPoint, Canva or Figma. Keep the honest framing: it complements Figma and Canva rather than replacing them, it's a preview, and it can burn usage quickly — so treat it as a fast way to explore directions, then polish in your real tool. The design demo turns a brief into a styled canvas you can tweak.

Chapter Six · Facts checked June 2026

Trust & Governance

For a business audience, capability is only half the decision. The other half is "can I defend this to risk, legal and the board?" This is the chapter that wins procurement — and the honesty in it is what makes the rest credible.

The three questions every business asks

Almost every stalled AI pilot dies on one of three questions. "Is our data used to train the model?" — no, not on business plans by default: Team, Enterprise and the API are governed by Commercial Terms that exclude them from consumer training, so for sensitive work you use that login. "Is it certified?" — SOC 2 Type I & II, ISO 27001:2022 and ISO/IEC 42001:2023 (AI management systems), with HIPAA-readiness via a BAA on sales-assisted Enterprise and GDPR support through a DPA with Standard Contractual Clauses. "Is it safe and well-governed?" — the ad-free pledge (no advertiser-influenced answers), the Responsible Scaling Policy with active ASL-3 safeguards, and Constitutional AI, all backed by published system cards and red-team results. The Governance Q&A gives you the answer for each stakeholder.

The honest limits

Teaching the limits isn't hedging — it's how you earn the right to recommend the tool, because a sceptical CFO trusts the person who names the risks. Five matter: hallucination (verify what matters, ground with sources); prompt injection in browser and custom-connector use (better, not zero — keep agents off banking and passwords); human-in-the-loop on any autonomous or scheduled run touching outbound actions or production systems (outputs as drafts until you've watched several runs); vendor and regulatory risk (frontier access can be suspended — keep fallback models, as Fable 5 taught us); and benchmark claims, which are often first-party, so pilot before you trust a superlative.

The Governance Brief

The deliverable that turns "I tried a cool AI thing" into "here's a governed plan" is a single page: the plan you'll use and why, the data-training stance, the certifications that apply, who needs to approve, and the guardrails you'll mandate — drafts-not-send, no sensitive-tab browsing, human approval thresholds. It's the document you forward to IT and procurement to get to "yes." The Governance Brief builder assembles a procurement-ready version as you tick the options. Governance isn't the boring slide — it's the slide that gets you the budget.

Chapter Seven

Your 30-Day Playbook

No new features here — just synthesis and commitment. The honest truth about any three-hour intensive is that you'll remember perhaps 30% of the demos, but you'll keep 100% of the plan you write down. So this chapter converts capability into a staged plan you'll actually follow.

Four weeks, advisor to executor

The arc mirrors the book. Week 1 — get fluent: bring one real task to Claude every day, default to Sonnet, keep adding to your prompt library, and turn on Memory and one safe connector. Week 2 — make it yours: build a custom Style from your real writing, set up your first Project, and use file creation on the spreadsheet or deck you most dread. Week 3 — delegate: run one real multi-step job in Cowork, try the Office add-ins on a live file or set up a scheduled task as a draft, and watch the first runs before trusting anything unattended. Week 4 — make it permanent and defensible: turn your most-repeated task into a Skill, complete your Governance Brief, pick your next course, and share one workflow with a colleague — teaching it is how it sticks. The 30-Day Playbook builder captures all of this as a one-page plan that saves on your device.

Which plan, and when to change it

Match the plan to the use. Start individuals on Pro (around $20/month) — it includes Opus, Projects, Artifacts, file creation, connectors and Cowork. Move to Max if you hit limits multiple times a week (the 5× tier is plenty for most). Move a team to Team — Premium seats for the Cowork and Code power users — and to Enterprise when you need SSO and SCIM, audit logs, the Compliance API, customer-managed keys, US-only inference, or a HIPAA BAA. The thresholds, not the brochure, should drive the upgrade.

Chapter Eight · Case study, all figures real

Case Study — Doing the Books with Claude

We finish where business audiences are most persuaded: not by a demo, but by a real company's high-stakes admin actually getting done. This chapter is a true, end-to-end account of running the books for Film Me Pty Ltd — a GST-registered Australian company — with Claude Cowork orchestrating and Claude in Chrome driving QuickBooks Online. Every figure below is real, including the moments Claude correctly stopped.

The setup, and the golden rule

The tool stack is Module 3 doing real work: Cowork orchestrates the multi-step job on the desktop, Claude in Chrome drives the QuickBooks UI directly (no API needed), a Python sandbox does the heavy data crunching, and the Gmail connector pulls receipts and remittance advices. One instruction shaped everything: "Prior financial years are already lodged with the ATO — do not touch them." Watching Claude respect that guardrail, again and again, is the whole lesson. Set your guardrails before the agent starts, not after.

Invoicing, reconciliation, and a $13,183.50 mystery

Three jobs, three lessons. Invoicing from a purchase order (the full account): Claude read a client PO from Gmail, duplicated the last good QuickBooks invoice, mapped every detail across, and caught the GST error a tired human misses — a second line that came in with no tax code — then verified the totals to the cent and stopped, handing back for a human to send. It then wrote the whole procedure up as a reusable Skill. Reconciling a year of multi-currency books (the full account): every account reconciled to its external statement to the cent, including a Wise AUD account that started $15,423 adrift because the bank feed had silently stopped for five months, dropping 304 transactions — recovered safely by diffing the statement against the register in Python so nothing was double-counted. GST, BAS and the tax hand-off (the full account): a year of GST mis-coding fixed, the $13,183.50 mystery customer payment untangled, and the headline finding that the BAS read $7,296 when the genuine current-quarter GST was only about $221 — the rest prior-period amounts that may already have been lodged.

The best move was to stop

And here is the real lesson of the entire course, made concrete. The smartest thing the agent did all day was to stop. Three times — the Wise USD opening balance, the prior-year correcting journals, and the BAS itself — Claude reached a point where the right action touched lodged returns or needed genuine professional judgement, and it handed back instead of guessing. It prepared the BAS so the numbers were visible, then stopped at the lodge step; it drafted the accountant email but didn't send it. That restraint — not the capability — is what makes the tool trustworthy for finance. The model worth teaching isn't "AI does your tax unsupervised." It's AI does the 90% that's tedious and error-prone, surfaces the 10% that needs a human, and leaves a clean trail for both. Claude is not a registered tax agent, and that is exactly the point.