Skip to main content

AI Assistant

Kara answers from the data in your own account - notes, transcripts, files in a space, what you have told it about yourself - not from general knowledge on the internet. That is the thing most worth remembering about this assistant: ask an ordinary AI assistant and you get back general understanding; ask Kara and you get back what actually happened in your meetings.

It is with you at both ends of a piece of work:

  • Thinking the strategy through with you - pulling together data spread over many meetings, going back over what mattered, comparing business figures, finding the thing that is stuck and needs solving.
  • Carrying the work out with you - sending email, building a report, a deck or a comparison table, adding a task right on the note.

Almost everything you ask Kara goes through the chat panel - it is where you ask, pick a skill, attach a file, and approve what Kara is about to do. The panel opens from any screen, and three shortcuts open it for three different situations:

Open withWhat you get
⌘ KA quick question. The cursor jumps into the Ask Kara box already in the bottom bar - you can type immediately, no panel opens and where you are working does not change. On a screen without that box, this key opens the chat panel.
⌘ ⇧ KThe whole chat panel. The same key opens and closes it - with the panel pinned, this is the only way out from the keyboard.
⌘ 3The layout with chat. Content + sidebar + a chat panel pinned alongside, full height. It sets the layout rather than toggling, so pressing twice changes nothing.

The first time you open it​

Kara shows the "Before you ask Kara" screen, which says two things:

  • To answer, Kara reads your notes, recordings and account information - so an answer is always grounded in your own data.
  • The content relevant to your question is sent to a third-party AI to produce the answer. This screen names which one (currently Anthropic Claude), with links to its Terms and Privacy Policy.

Click I agree, ask Kara to use it - it is the only button; there is no decline, closing the chat panel is the way out. Asked once, remembered on this device.

The Before you ask Kara screen inside the chat panel: an explanation of what data Kara reads, a "Where the data goes" block naming Anthropic Claude with Terms and Policy links, and an "I agree, ask Kara" button at the foot

What Kara reads​

Six sources, and knowing them is knowing why the same question is sometimes answerable and sometimes not:

SourceHolds
Notes and transcriptsYour whole note history - content, summaries, tasks, participants.
Files you uploadedFiles in a space, and files you attached to the question itself.
What you told it about youWhat should Kara call you? and Instructions for Kara under Settings → General.
Organization instructionsIf you belong to an organization - see Organization settings → General.
External systemsData from the connectors you have linked and left switched on.
The webOnly when you ask for it, through the Search the web for skill. See Skills.
Reading requires permission, and some data passes through a partner

Kara reads only the sources you have agreed it may: the Before you ask Kara screen opens the gate for notes and account data, and each connector has its own consent gate when you link it. To produce an answer, the content relevant to your question is sent to a third-party AI for processing and the result comes back. What exactly you agreed to, and how to withdraw it, is in Consent & permissions.

What to ask​

Kara suggests things based on where you are standing. Inside a note: Summarize this note, Pull out the tasks, Draft a follow-up email. On the home page: What should I focus on today?, Summarize my recent meetings.

But a suggestion is only a starting point. The genuinely valuable questions are usually more specific:

  • "What did Tuan dislike about the 12th-floor unit?"
  • "Which three clients did I promise a quote to this week?"
  • "Since the start of the month, what have clients asked about most?"
  • "Compare what the two sides agreed last time with what they agreed today."

Asking about one particular passage​

Select any passage in a note, a summary or a transcript and choose Ask Kara about the selection. That passage is attached to the question, so Kara knows exactly what you mean - no describing it again in words.

Asking Kara to do, not just answer​

Kara can do work right on the note: drafting and sending email, building a report, a deck or a comparison table, adding a task.

Before writing anything into the note, Kara asks "Create this on the note?" and you either create it or choose Later. If there is something to think about, Kara adds "Worth a look first" - read that part before you agree.

Work that leaves Kara gets one more layer: an email only leaves the machine through the mail connector you linked, and always after you confirm. See Sending email and Connectors.

Focus on - what Kara reads first​

The same question gives different answers depending on what Kara reads first. That is shown right above the composer, as a small chip: Focus on.

Kara works this chip out from where you are standing - it is not something you pick from a list. Open a note and the chip carries that note's name; stand in a space and it carries the space's name; open someone's profile and it carries their name.

The chip saysKara reads firstFits
A noteThe open meeting"What worried the client most?"
A space · N notesThe notes in that space, plus its files"What is still blocking the Sunrise project?"
A contactEvery meeting with that person"How many times has Minh changed his mind about the budget?"
Calendar · Todos · Contacts (the address book) · ArtifactsExactly the screen you have open"Which tasks are overdue?"
Your dayThe home page - your today"What should I focus on today?"

A priority, not a filter​

This is the easiest thing to misread. Focus on does not cut down what Kara may read - it only sets the order. The chip says as much when you hover it: "What Kara is focused on. Everything else is still available, just lower priority."

Kara answers from the highest-priority place first, then widens on its own when it needs to - when what it is reading does not carry enough grounds, or when the question mentions something outside it. You do not have to press anything to widen the scope.

Clearing Focus on​

With a note open but wanting to ask something else entirely - "how is commission usually split?" - click the Focus on chip: bring the cursor onto it, a ✕ replaces the icon, and clicking it is all it takes. The next question runs across the whole account without carrying the open note along; the conversation inserts a divider line noting that the subject changed.

Three things worth knowing about clearing it:

  • It only holds while you stay where you are. Moving to another note, another space, or back to the home page is a fresh detection - the chip returns.
  • On the home page there is no chip. "Your day" is not somewhere you deliberately navigated into; Kara still prioritizes your today, there is simply nothing to clear, so no chip is drawn.
  • The "All notes" chip has no ✕ - there is nowhere wider to step out to.
Getting "not found"?

Before concluding Kara does not know, look at the Focus on chip. Asking "which clients are waiting on a quote" while a note is open makes Kara read that meeting first - it will usually widen anyway, but clearing the chip and asking again is the fastest way to be sure.

Even at the widest scope, Kara only reads what your account has permission to see. A colleague's note that was never shared with you is in no scope at all.

Reference files in a space​

A space can hold files of its own - price lists, sales policy, project descriptions. Ask while Focus on is that space and Kara reads those files too, not just the meetings.

This is how questions the meeting itself never covered get answered: "Is there still a discount available on this unit?" - if the price list is in the space. See Spaces.

The composer: the + button, / and @​

Beyond typing a question, the composer has three shortcuts: + attaches something, / picks what Kara will do, @ points at what the question is about. All three attach to one turn, appear as chips above the input, and can be removed.

The + button​

Three items:

ItemDoes
Select filesAttaches a file from your machine to the question.
SkillsThe list of jobs you have taught Kara - the same list / opens, for anyone who wants to see what is there before choosing.
ConnectorsSwitches individual connectors on or off for this chat.

Switching connectors from the chat panel​

+ → Connectors lists the systems you have linked that Kara is currently allowed to consult. Switch one off and Kara stops looking there; switch it back on and it is available again.

Switching off here is not disconnecting

This switch is a preference for the chat panel only - it does not revoke the access you granted. To disconnect properly, go to Settings → Connectors. A newly linked connector starts on, so you do not have to find this menu before you can use it.

At the foot of the list is Manage connectors, which opens the Connectors tab in Settings.

Type / to run a skill​

Type / at the start of the composer to open Run a skill and pick one by name. The chosen skill becomes a chip in the input, and the input's placeholder changes to whatever that skill needs to know.

Kara also picks a skill itself when your question is obviously that job. When it does, the chip is marked as Kara's choice - remove the chip to ask as normal. With no skills switched on, the list says so and points the way: No skills enabled yet - add them in Settings → Skills. See Skills and Agents.

Type @ to point at exactly the right thing​

Type @ to open Mention something. Six kinds, each meaning something different when the question is sent:

MentionWhat Kara does
A contactLoads what Kara holds about that person.
A noteLoads that note's summary and transcript, even when it is not open on screen.
A spaceNames the notes in that space.
A connectorYou are pointing at a source to consult: "check @HubSpot and see". A connector not yet linked shows Not connected on its row - Kara says plainly that it cannot check, rather than pretending it did.
A todoLoads that task, who has it, when it is due, and which note it came from.
An agentRuns this whole turn in that agent's role.

While Kara answers​

The answer appears word by word. Above it is a line telling you Kara is working, and how much you see depends on what Kara is doing:

  • Answering a question - the line "Kara is thinking…" pulses, then the answer appears. It is a sign that work is happening, not a transcript of the reasoning.
  • Producing an output (slides, a diagram, an email, a mind map) - Kara shows the whole sequence: a Creating … line with Step 3/5, and View steps expands all five: Reading the note → Need to ask you → Waiting for your review → Saving to the note → Summary. Each step has a line saying what is happening, so you know whether Kara is reading, waiting on you, or finished writing.

The sequence is collapsed by default - it sits right above the composer, and five open lines would push what Kara just made off the screen.

If Kara is running long or heading the wrong way, press Stop.

The three buttons after every answer​

Make it shorter · More detail · Change the format - faster than retyping the request.

Timestamps and the copy button​

Under every message - yours and Kara's - there is a timestamp and a copy button. On a computer they appear when you bring the cursor onto the message; on a phone they are always there, because a tap has nothing to hover with.

  • A message from today shows hh:mm; anything older adds the date - a conversation lives across several days, so "14:32" alone would be confidently wrong. Hover the number for the full timestamp.
  • The copy button takes the content as you read it. Once copied, the button turns into a tick for a beat.
  • Messages from before Kara started recording times show no time at all - better none than an invented one.

Where the chat panel sits​

  • Pin to the right edge - reading the note and asking at the same time, good on a wide screen.
  • Switch to a floating panel - a panel floating mid-screen, good on a small laptop.

On a computer, ⌘ 3 opens the layout that already includes the chat panel.

How long one conversation lasts​

By default you stay in the same conversation until you start a new one yourself. It is not cut short by moving to another note, by finishing a recording, or by reloading the page - changing screen only changes the Focus on chip and inserts a divider; the thread runs on.

Exactly three things start a new conversation:

EventWhat happens
You click New chatThe open conversation is kept in history; you start from a blank page.
You sign outThe next sign-in opens a new conversation. The old ones are still in history.
You leave the app for over 2 hoursClosing the tab, killing the app on your phone, or simply leaving it - counted from the last thing you did in that conversation. Come back sooner and you carry straight on.

Chat history​

Every conversation with Kara is kept. You can go back and read one, or carry on asking without rebuilding the context from scratch.

In the chat panel, click Chat history. The list is ordered most recent first, each row carrying the title Kara gave it, how long since the last message, and the message count.

  • Search conversations - the box at the head of the list. It searches both titles and message content, because what people remember about an old conversation is usually something said in it, not the name it was given. With no matching row, the list says No conversations match - which is different from No chat history yet.
  • Delete a conversation - the bin icon at the end of the row, with a confirmation step. Deleting cannot be undone, and Kara stops using that conversation as grounds for answers from then on.

Within a conversation, Kara remembers what was said in earlier turns. That is what lets you ask short questions:

  • "Summarize yesterday's meeting with Tuan."
  • "What was he most worried about?"
  • "Draft me an email answering exactly that."

Open a new conversation and Kara does not carry that thread over; you have to set it up again.

When to start a new one​

A real change of subject deserves a fresh start. One conversation dragging across five different clients makes it harder for Kara to hold your meaning, and harder for you to find again later.

A simple rule: one job, one conversation.

If you remember asking something but not where, use search - the results include conversation content, not just notes.

Checking an answer​

Under every chat panel there is a small line: "Kara can make mistakes." It is not boilerplate hedging - it is how you use Kara safely.

Three principles run underneath every answer:

  1. Answer only from your data - notes, transcripts, files in a space, the information in your account.
  2. No grounds means saying so - rather than producing a plausible-sounding number.
  3. Never write out on its own - every write has to be confirmed by you first.

There is still room for error: a transcript mishearing a number, a joke read as a commitment, two similar voices attributed to the wrong person.

A ten-second check​

For any answer you are about to act on:

  1. Click the quote - read the original passage Kara leaned on. See Quotes.
  2. Cross-check the numbers - price, floor area, dates, percentages.
  3. Check the people - who said it, who took the task.

Three signs to stop and check​

  • The answer carries no quote where it clearly should.
  • A number is suspiciously round where the meeting was vague.
  • Kara asserts something nobody mentioned in that meeting.

The fastest self-check is to ask back: "Which passage is this based on?", "Did the client actually say that number, or did I infer it?", "If you are not sure, say you are not sure."

Where to be most careful​

TaskWhy
Sending a price to a clientOne wrong number costs credibility, sometimes money.
Committing to a dateKara records what you said, not what you meant.
Writing data into another systemAn error at the source spreads into every report downstream.
Forwarding a summary as-isIt may carry remarks not meant for outside eyes.

Reporting a bad answer​

When an answer is clearly wrong, send feedback: account menu → Help → Send feedback → choose Something is broken. Describe what you asked and what Kara answered - it goes straight to the team.