Clickstarter Education · Interactive guide
A working guide for Executive Assistants. Everything demonstrated in the session, plus the prompts, the limits nobody mentions, and the guardrails that keep you and your executive safe.
Two thirds of Australian workers already use an AI tool at work. Under one in ten has it actually built into a workflow. The gap between playing with Claude and running work through Claude is where the whole advantage sits.
The 8% comes from a 2026 survey of 600 administrative and executive assistants. The sample is 77% United States and only 4.3% Australia 50. Treat it as indicative, not local. In the same survey ChatGPT sat at 70% and Microsoft Copilot at 43%. The point stands either way: this room is early.
Same survey, same caveat. Useful because it matches what the job feels like.
Source: Vimcal and CCing My EA 2026 Report, n=600 50. Separately, Australian office workers report 47% of work time going to low-value repetitive tasks against a 41% global average, though that figure is from 2024 49.
Cowork is the same agentic engine that powers Claude Code, without a terminal 1. Anthropic's own line: describe a desired outcome, then return later to completed work 1. It is the difference between asking a question and delegating a job.
Anthropic analysed 1.2 million anonymised Cowork sessions during May 2026 across more than 600,000 organisations. Business process and operations work was 33.4%. Content and copywriting 16.4%. Software development only 8.7% 1314. The most common use of the agentic product is not coding. It is the kind of work that lands on your desk.
Each session gets an isolated temporary sandbox on Anthropic servers, created at start and destroyed at end, sharing no state with other sessions or organisations 4.
Cloud sandboxes enforce a mandatory egress proxy with allow-listed destinations and session tokens that expire within hours 4.
When a cloud session needs a local file it travels through an Anthropic-brokered connection via the desktop app, limited to folders you connected, checked against your permissions per call 4.
Close the desktop app and cloud sessions keep running. They just lose access to your local folders, local connectors, the browser and computer use 24.
The agent loop runs natively on the device, with code execution inside an isolated Linux virtual machine using the operating system hypervisor. Virtualization.framework on macOS, Hyper-V on Windows 4.
Live artifacts and plugins are desktop-only 2. Browser automation and computer use also need the app open 4.
| Item | Position as at August 2026 |
|---|---|
| Desktop, macOS and Windows | Generally available since 9 April 2026 968 |
| Web and mobile | Beta, rolling out from the Max plan since 7 July 2026 39 |
| Plans included | Pro, Max, Team, Enterprise. Free is excluded 211 |
| Price | Published in USD only. Pro US$17 per month annual, Max from US$100, Team US$20 per seat annual 11. No official AUD price |
| Australian data residency | Listed as "Coming 2026" 12 |
| Session sharing | Cowork sessions cannot be shared with other people 2 |
| Memory | What Claude remembers in chat does not carry into Cowork sessions yet 2 |
Do not start with something precious. Start with something tedious and reversible.
I want you to work through this without checking in on every step. Task: go through the folder I have connected and produce a single summary document. For every file in the folder: 1. Record the filename, the date it was last modified, and one sentence on what it contains. 2. Flag anything that looks like a duplicate or an outdated version. 3. Flag anything containing personal information that should not sit in a shared folder. Output: one Word document called Folder Audit, with a table for point 1 and two short lists for points 2 and 3. Do not delete or move anything. Do not rename anything. Tell me at the end what you were unsure about.
Anthropic has not published a quantified time-saved figure for Cowork. The named case study is Thomson Reuters, whose CTO says the human role becomes validation, refinement and decision-making rather than repetitive rework 16. No hours figure is given. If you see a percentage on a vendor blog, treat it as marketing.
This page is the example. A single HTML file, your colours, your fonts, your logo, every claim numbered and traceable. Built in one session and hosted anywhere.
Inside the artifact viewer a strict content security policy blocks external scripts, fonts and images, so Claude inlines everything 31. There is no backend, so a static page cannot store form submissions or authenticate viewers 31. Rendered size limit is 16 MiB 31. Persistent artifact storage on paid plans caps at 20 MB, text only, and only works once published 30.
In July 2026 public Claude share links were indexed by Google, exposing medical records, employee reviews and internal documents 33. On Pro and Max a public link is the only sharing option. Team and Enterprise can share with named people 32. Audit Settings, then Privacy, then Shared Chats.
The Claude Opus 5 System Card, dated 24 July 2026, states that the model "hallucinates factual claims slightly more than Opus 4.8, despite being more accurate overall". It also records "a surprising number of cases in which Opus 5 confidently stated an answer about which it was in fact unsure" 41.
Better model does not mean less checking. It means the wrong answers arrive with better posture. For an EA, whose credibility is the product, that is the whole ball game.
Go back through what you just wrote. For every factual claim, every number and every date: 1. Give me the exact source URL. 2. Quote the sentence from that source that supports it, word for word. 3. If you cannot produce a quote, delete the claim and tell me you deleted it. Then list separately anything where sources disagreed with each other, and tell me which one you used and why. Do not defend anything. I would rather have a shorter document that is right.
The honest version, because your executive will hold you to whatever you promise.
Claude does the research, holds every constraint at once, and produces the itinerary document. From an Australian account that means web search rather than the Expedia connector, which is United States only for now 71. A human confirms the fare and books it. That is still the win, because the slow part of executive travel was never the click that buys the ticket. It is reconciling a departure window against three meetings, a time zone, a loyalty preference, a budget policy and a return that lands before school pickup.
Fill this in and copy the result. It produces a prompt that holds every constraint at once, which is the part humans lose track of.
Claude creates .docx, .pptx, .xlsx and .pdf on all plans, across web, desktop and mobile, capped at 30 MB per file 8. Purpose-built skills exist for each format 35. The sandbox that builds them runs Python with internet access disabled, which is the documented reason file creation is comparatively safe 69.
Audience, purpose, length, tone, and what happens after they read it. Anthropic's golden rule: if a colleague with minimal context would be confused by your prompt, Claude will be too 57.
Say the file type, the section structure and the word count. Opus 5 writes long by default and expands scope, so constrain both explicitly 58.
Three to five examples beats any amount of description. Put long reference documents first and your instruction last, which Anthropic claims lifts quality by up to 30% 57.
Opus 5 self-corrects, so legacy instructions like "include a final verification step" now cause over-verification and waste 58. Modern models need less scaffolding, not more. The best prompt achieves the goal with the minimum necessary structure 59.
This is how the workshop you attended was built. One project, holding the whole sequence as instructions, so every workshop starts from a standard instead of a blank prompt.
Title, content description, target audience. Three questions, asked in order, every time.
A cited research report saved to Drive before anything is designed. Facts first, format second.
Event description, interactive guide, interactive slide deck. All from the same researched material, so the three agree with each other.
You are helping me produce [deliverable type] for [organisation]. Always follow this sequence and never skip a step unless I say so: 1. Ask me for the title. 2. Ask me for the content and purpose. 3. Ask me who the audience is. 4. Research the topic and give me a cited summary before producing anything. 5. Produce the deliverable. House rules for every output: - Colours: [your hex codes, with a note on which is background and which is text] - Fonts: [heading font and weight, body font and weight] - Tone: [plain, direct, no jargon] - Australian English spelling throughout. - Every statistic carries a numbered citation and a working source URL. - If you cannot verify a claim, say so rather than smoothing over it.
| Connector | What it does for this work |
|---|---|
| Canva 21 | Search, create, autofill and export designs from a prompt. Brand-template autofill, resize and export. Ships a Branded Presentation skill. |
| Google Drive 17 | Reads Docs, Sheets, Slides, PDFs and Office files. Writes by uploading files, creating folders and saving Claude-generated files. Text only, embedded images are not processed. |
| Asana 22 | Search, create, update and track tasks, projects and goals. Can turn a conversation into a project with owners and timelines. |
| Zoom 20 | Search meetings and chat, retrieve AI summaries, transcripts and next steps, create Zoom Canvas docs. Needs Workplace Pro, admin rights, Smart Recording and Meeting Summary switched on. |
| Slack 19 | Search messages, channels, threads and files. Draft messages for review, create canvases, schedule messages for later. |
Cowork scheduled tasks run hourly, daily, weekly, weekdays only or on manual trigger, on all paid plans 7. They run remotely, so they fire even when your computer is asleep and the app is closed 7. One exception: a task needing local files runs locally and needs an awake machine 7.
Every firing runs as its own fresh session with no memory of prior runs 7. Your brief cannot say "unlike yesterday" unless it re-derives yesterday from a source it can read. Build it stateless.
Microsoft 365 with write consent can draft and send email 18. The Gmail connector creates drafts only and, in Anthropic's words, cannot send emails on your behalf 17. Know which one you are on before you promise anything.
Tick what you actually have connected. The prompt on the right updates as you go. Copy it into a scheduled Cowork task.
Third-party commentary, not confirmed by Anthropic, warns that scheduled runs consume materially more tokens than chat and that a few daily tasks can exhaust weekly limits on an entry plan. The same source notes timezone handling, retry behaviour and notifications are undocumented 70. Treat those as open questions. Start with one task, watch it for a fortnight.
You handle the most sensitive material in the building. Tap each card for the fix.
Privacy Act tranche two is not enacted and has no fixed commencement date. Removal of the small business exemption is a proposal, not law 55. Separately, a "Deepfake Defamation Act 2026" appearing on some Australian legal marketing sites has no bill number and no parliamentary reference. Do not cite either.
Working in government? DTA guidance is that OFFICIAL: Sensitive or above must not go into public generative AI tools, personal information likewise, and human judgment must always guide how the tools are applied 52.
Tick as you go. Nothing here is stored, so finish it in one sitting or screenshot it.
Nothing ticked yet. Start with week one. It takes about fifteen minutes and everything else depends on it.
Copy, paste, edit the bracketed parts. These are written the way Anthropic recommends: specific about format, honest about constraints, with explicit permission to say "I don't know" 4057.
Prepare a one-page brief on [name] at [organisation] ahead of a meeting on [date]. Cover: current role and how long they have held it, their previous two roles, anything they have said publicly in the last twelve months, their organisation's recent news, and any connection to us that you can find. Rules: - Every claim gets a source URL. - If you cannot verify something, leave it out and list it separately as "could not confirm". - No speculation about their personality or intentions. - Maximum 400 words plus the source list.
Here are the notes from [meeting]. Turn them into a decisions and actions document. Structure: 1. Decisions made, one line each, with who decided. 2. Actions, each with an owner and a due date. If an owner or date was not stated, write "unassigned" rather than guessing. 3. Open questions nobody answered. 4. Anything I should chase before the next meeting. Do not invent an action that was only implied. If it was vague, put it under open questions.
Below are three emails [name] wrote. Study the voice: sentence length, how they open, how they close, how direct they are, what they never say. [paste three emails] Now draft a reply to the email below in that voice. Keep it to [x] words. Then tell me, in two lines, which parts of the draft you are least confident about matching.
Review this before it goes out to [audience]. Check for: factual claims I cannot back up, numbers without a source, anything that commits us to something we have not agreed, tone that does not match the relationship, and anything that would embarrass us if forwarded. List problems only. Do not rewrite it. Rank them worst first. If you think it is fine, say so plainly rather than finding something to say.
Seventy sources, numbered to match the markers throughout this guide and the workshop deck. Checked August 2026. Links go stale. If one breaks, search the title.