Using the Chat
Talking to Your Agent.
The chat is where all the work happens. Reports, emails, research, files, follow-ups. Everything triggers from a plain message. No commands to memorise. No special syntax. Just tell your agent what you need, the way you'd tell a colleague.
How to Ask
Write the way you actually speak. Your agent is built to understand natural language, intent, and context. Not keywords. Don't over-specify. If it's missing something, it'll ask you for it before it acts.
Plain English works best.
"Summarise the Mercer contract and tell me if payment terms are standard" lands cleaner than "Execute doc_summarize contract=Mercer.pdf flag=payment_terms".
Don't front-load every detail.
Start with the outcome you want. It'll confirm what it's about to do and ask follow-ups if it needs more.
Conversations have memory.
Ask a follow-up, add context mid-thread, refine. It remembers the full conversation. You don't need to re-explain.
It Drafts First. You Approve.
Anything with external blast radius pauses for your sign-off. It writes the message, shows you, and waits. No "oops I already sent it." This isn't a feature you configure. It's how your agent is built.
Emails to real people
It drafts, shows you the recipient, subject, and body, and waits. You hit approve. It sends.
External messages
SMS, Slack to other companies, anything that leaves your tenant. Drafted first, sent on your word.
Payments & invoices
Anything that moves money or commits you financially pauses for your sign-off. No exceptions.
Calendar sends
Internal calendar blocks it'll just do. Invites that go to outside attendees get the same draft-first treatment.
The default is pause. If you ever want to ease a specific flow. Say, auto-send internal status updates. You tell it once and it remembers. But it never relaxes the default without you saying so.
When It Gets It Wrong
Correct it in the chat. Plain language. "That's not quite right. The client is Mercer, not Mercier." "Use my warm tone, not my formal one." "Don't loop David in on this thread."
Your agent logs the correction into its memory the moment you say it. Next time the same context shows up, it applies what you taught it. Corrections compound. Month one it's useful, month six it's trained to you.
Your agent won't argue with a correction. If you tell it it got a name, a preference, or a workflow detail wrong, it takes the correction on the chin and writes it down. It'll push back on something structural. "you asked me to skip approval on external emails, I won't do that": but never on something you know better than it does.
Attaching Files
Hit the paperclip icon in the chat input to upload. Your agent will read the file, extract what's relevant, and answer your question against its contents. Not against guesses.
Supported formats:
For images, it reads text inside them (OCR), identifies objects, and can describe what's in the frame. For spreadsheets, it reads the cells, not just the rendered table. For PDFs, it handles scanned documents too.
Multi-Step Requests
Your agent chains tools. One request can touch email, calendar, a document, a web search, and its memory. All in the same turn. You don't orchestrate it. You describe the outcome; it sequences the steps.
Example chain.
"Find an hour with Sarah next week, draft a brief for the meeting, and send her the invite."
- 1. Checks both calendars, finds an open hour.
- 2. Pulls context from your last three threads with Sarah.
- 3. Drafts a one-page brief and saves it to your Drive.
- 4. Writes the invite with the brief attached.
- 5. Pauses. Shows you everything. Waits for your approve.
Example Prompts
Real prompts that work. Copy these, tweak them, or just read them to get a feel for the register.
Check my inbox overnight and tell me what actually matters before my 9am.
Do I have anything Friday afternoon? If not, book a 30-min slot for the Mercer follow-up.
Read the contract I just dropped in and flag anything non-standard on payment terms.
Draft a reply to David's email from yesterday. Warm, but push back on the timeline.
Pull recent sales in Richmond VIC 3121, last 90 days, and summarise the price bands.
Find an hour next week with Sarah, draft a brief for the meeting, and send her the invite.
What It Won't Do
It refuses a narrow set of things, always, and tells you why when it does.
- Harmful content. It refuses and tells you why.
- Cross-tenant actions. It can't reach into another user's agent or data, and won't pretend to.
- Unauthorised external actions. Sending from an account you haven't connected, impersonating someone, acting on an identity that isn't yours.
- Unverified claims presented as fact. If it doesn't have the source, it says so. It won't make it up to sound useful.
Refusals are logged. Every time it says no, there's a reason written down. What was asked, why it declined, what would have to change for it to be allowed. You can review the log any time. Nothing disappears.