A fast, practical session for Executive Assistants. With Dante St James.
Hand Claude a multi-step job, walk away, come back to finished files.
Real documents for staff and clients, with every claim traceable.
What is real, what is marketing, and where the human stays in the loop.
One place, one standard, no rebuilding from scratch each time.
Calendar, inbox, chat and news, on your desk before you are.
What an Australian EA must not put into an AI tool, and why.
One word we use all day: a connector is just a link between Claude and an app you already use, like your calendar or your inbox. Three live exercises, so bring a laptop if you have one.
Almost nobody has done the second part.
Salesforce and YouGov, 1,293 Australian workers, fielded 18 to 22 May 2026 46
Vimcal and CCing My EA 2026 Report on the Administrative and Executive Assistant Profession, n=600 50
In the same survey ChatGPT sat at 70% and Microsoft Copilot at 43%. Claude was tied with Otter.ai 50.
Eighty-six per cent believe AI will enhance the role rather than replace it 50. That is the useful part. The 8% is the edge, and it will not last.
“Rather than responding to prompts sequentially, Claude tackles intricate, multi-step tasks autonomously. Describe a desired outcome, then return later to completed work” 1.
Start a job on your laptop and pick it up on your phone. Your files follow you 3. Big jobs get broken into pieces and worked on at the same time 1.
Available on all plans across web, desktop and mobile, capped at 30 MB per file 8. While it builds a file it is cut off from the internet, which is what makes this the safest thing it does 69.
On Mac and Windows it finished testing on 9 April 2026 968. In a browser and on your phone it is still being tested, and is arriving first for people on the dearest plan 3. Included on Pro, Max, Team and Enterprise. Not on Free 211.
1.2 million sessions with names stripped out, May 2026, across 600,000+ organisations 1314
A colleague cannot open the Cowork job you are running 2.
What Claude picks up about you in ordinary chat does not follow you into Cowork yet 2.
Anthropic lists Australian storage as coming in 2026, so it is not here yet 12.
Anthropic has published no hours-saved figure. Treat any percentage you read online as advertising.
Thomson Reuters tested Cowork against their real Microsoft 365 setup. Their CTO: “The human role becomes validation, refinement, and decision-making. Not repetitive rework” 16. No hours figure is given, and that is worth saying out loud.
Something tedious and reversible. Never something precious.
A project keeps its own chats, its own files and its own rules 34.
Hex codes, fonts, tone and the build sequence, written down and reused.
Two chats in the same project do not share what you told each other. Only the saved files do 34.
A set of written instructions Claude reads whenever the job calls for it 35.
It stays out of the way and loads only when the job matches it 35.
Your organisation can switch one on for everybody at once 61.
How it sounds, not how it looks. Presets plus custom 62.
Upload writing your executive actually wrote 62.
Your prompt beats the design system, which beats Claude's own taste 31.
The Claude Opus 5 System Card, 24 July 2026: the model “hallucinates factual claims slightly more than Opus 4.8, despite being more accurate overall” 41. In plain terms: it is right more often, and it also makes things up a little more often.
The same card records “a surprising number of cases in which Opus 5 confidently stated an answer about which it was in fact unsure” 41.
The wrong answers now arrive with better posture. For an EA, credibility is the product.
Their guidance: these techniques reduce hallucinations but do not eliminate them. Always validate critical information 40.
External context: eight AI search tools were collectively wrong on more than 60% of queries in a 2025 Columbia Journalism Review test, with paid tiers more confidently wrong 42. Older models, same lesson.
Anthropic's documented technique: explicitly allow Claude to say it does not know, rather than leaving it to guess 40.
For long documents, have it extract word-for-word quotes before answering, then retract any claim without supporting text 40.
Web search responses carry citations automatically, with up to 150 characters of cited text 25. Anthropic offers citations as the way to check, not as an accuracy promise 25.
What still works from Australia: Claude searches the web for fares and rates, holds every constraint at once, and writes the itinerary. You confirm and book. The slow part of executive travel was never the click that buys the ticket.
Title, content, audience. Three questions, asked in the same order every time.
A cited report before anything is designed. Facts first, format second.
Description, guide and deck from the same material, so all three agree.
Makes designs and drops your details into templates you already use 21.
Reads your documents and saves new ones back. It reads words, not pictures 17.
Turn a conversation into a project with owners and timelines 22.
Meeting notes, summaries and next steps. Your IT admin has to switch it on first 20.
“Show your prompt to a colleague with minimal context and ask them to follow it. If they'd be confused, Claude will be too” 57.
Anthropic claims up to a 30% quality improvement from putting long documents first with the query at the end 57. Three to five examples beats any amount of description.
Claude now checks its own work, so telling it to “double check at the end” just makes it slower 58. It also writes long and does more than you asked, so say how long and where to stop 58.
You can set a job to run every hour, every day, every weekday or every week, on any paid plan, and it runs with your laptop shut 7. But every run starts from nothing 7.
It cannot say “unlike yesterday” unless it re-derives yesterday from a source it can read.
On Microsoft 365, once your IT admin allows it, Claude can draft and send 18. Gmail creates drafts and, in Anthropic's words, cannot send on your behalf 17.
Nothing ticked. Start with week one. Fifteen minutes, and everything else depends on it.
Confidential exec material belongs on a Team or Enterprise seat, not a personal account 53. The OAIC says do not put personal information into publicly available AI tools 51.
Everything from today, plus the prompts, the limits and the guardrails. Working builders you can use tonight.
Scan it, or type it:
t.ly/-Ts0k
Seventy cited sources, checked August 2026. Products change fast. Verify anything critical against the source before you rely on it.