Everything she can do,
and what she asks first

One live audio socket carries the conversation: she listens and speaks natively and can call a tool mid-sentence, so a simple answer arrives in under a second instead of the fourteen a transcribe-think-speak chain used to take.

Say her name to wake her; she sleeps again on silence. Talk over her and she stops at once and follows the new instruction, barge-in is built into the protocol, not bolted on.

Anything that needs real thinking is handed to Claude in the background while she keeps talking to you. You hear "one moment" and then the answer, not a spinner.

“Justine, actually, forget the mail. What did the printer say about the paper?”

She reads unread mail, watches for people you know, and sends. Names are resolved against a contact book she keeps: a misheard "Pria" never becomes a message to the wrong Priya, because when the match is not certain she asks instead of guessing.

WhatsApp goes through the web client in your own Chrome, text, a screenshot she just took, a file found anywhere on the Mac by the name you call it, a reel you were watching. Gmail is read and sent natively; with the optional Gmail API connection she can also label threads and put drafts into your Drafts folder.

Sending always waits for a yes. Reading, drafting and finding never do.

“Send Priya the launch deck and the analysis report on WhatsApp, and tell her the invoice is coming Friday.”

Google Calendar, read out loud on request or as part of the morning. She blocks focus time, frees it again, and moves things when you say so.

Reminders are spoken. If you are away from the Mac they can ring your phone instead, so "remind me to call Dad at seven" means she will actually find you at seven.

Timed work is a schedule she keeps: "at five, send me the summary of today" is one sentence and one entry.

“Block me till noon, move the printer to Thursday, and what is after lunch?”

She controls the Chrome you already use, your logins, your tabs, through Chrome's own scripting. There is no second browser to sign into and nothing to sync.

Short errands use a reflex loop with no language model in it: she maps the page a few screens deep, numbers every control, decides the next click with a small typed decision model, acts, and checks. "Search Google for X", "open YouTube and play X", "open my GitHub notifications" measure three to six seconds. Anything longer or fuzzier goes step by step with Claude reading the page.

She can read a page back to you cheaply, or look at it as a picture when layout matters. Destructive controls, delete, send, pay, are walls she does not cross without asking.

“Open Pinterest, search matte packaging, save the second one to the packaging board.”

Opens any app by name. Lists what is in front. Takes one deliberate screenshot when you ask her to look, never on a timer, never kept, and describes it or acts on it.

Real mouse and keyboard through Accessibility, so "click the blue button" and "type the address in" work in apps that have no scripting of their own. Volume, music, the terminal, the editor.

Finds any file or folder on the Mac by the name you called it, wherever it is, and makes it sendable.

“Look at my screen, what is this dialog asking? Say yes to it.”

Presentations. A deck from a sentence or from a brief in your files, sections, slides, your palette, shown on a surface over the orb, exported to Keynote or PDF.

Image studio. She art-directs before she generates: two considered prompts, the better one shown, sent wherever you say.

The Forge. A 3D model, a scene, or a simulation you can turn with the mouse, written as one Three.js program and rendered live.

Blender. Packaging and signage mockups rendered from a description, when Blender is installed.

Sketch and analyse. A quick drawing; a document read and answered with the source named, a PDF, a spreadsheet, a photo of a whiteboard.

“Make a nine-slide deck from yesterday's brief in the brand palette, then export it to Keynote.”

Say what you want built. She brainstorms it with you out loud, one question at a time. When it is settled she writes the design document into the project, then starts a real Claude Code session and drives it through plan, review and phased execution, several sessions, if it is big.

Every run is recorded: the prompt, the model, the transcript, the tokens, the outcome. Nothing spawns Claude Code outside that pipeline, and a run always ends in a terminal state, it cannot be left "running" forever.

She watches every Claude Code session on the machine, not only hers. When one is blocked on a permission prompt she says which one and what it asked; "yes" answers it. You can steer a running session by voice, read a repository without spending a model call, open the project in VS Code or Terminal, and ask how the subscription's five-hour and seven-day windows are doing.

“Build me a gallery page for the portfolio., Which one: masonry or grid?, Grid., Lightbox on click?, Yes., Writing the spec.”

Say goodnight and name an hour. Until then she sweeps the inbox every ten minutes, files each new thread under a label, client, vendor, newsletter, needs reply, urgent, personal, and drafts the replies that need one. Drafts go to your Drafts folder and are never sent.

Anything angry, and anything about money, billing, invoices, contracts or legal, is flagged for you rather than answered. Every action is an event in the night's file; the morning report is written from those events, so it can only say what actually happened.

At the hour she rings your phone: "Now or in five minutes?" Now brings the report as her voice and as text, with "approve all drafts", which sends them, or "call again at eight". The Mac stays quiet until you say.

“Goodnight, Justine. Call me at six with the report.”

What you ask her to remember lives as plain Markdown in a folder on your Mac. Open it, edit it, delete a line, she is not a database and does not pretend to be.

Your own files are her knowledge: drop a PDF, a spreadsheet, a Word file, a photo of a whiteboard on the Knowledge tab and it becomes a page she answers from, with the source named. Facts about people, a client's name, a school, an award, are answered from those files, never carried in her head from a persona.

Everything said is kept and searchable on your disk: "what did we talk about yesterday" is answered as a day, not a word match.

“Remember that Priya prefers WhatsApp and never the old teal.”

Pair a Telegram bot and she reaches you away from the Mac: the overnight report, a question that cannot wait, buttons to approve. Give her a second Telegram account and she rings you with a real call in her own voice, free.

If you want cellular calls, Twilio is supported; it is the one paid option and it is off unless you set it up.

“Call me when the build finishes.”

Illustrator, Photoshop and InDesign are scripted directly: open a file, export the active artboard to PNG, JPG, PDF or SVG at any scale, save. Canva and Figma through their MCP servers; Notion for the board and the briefs. Pinterest, Behance, Dribbble, Google and Adobe Fonts, Coolors, Unsplash and Pexels in your Chrome.

"What size is a LinkedIn banner", "how much bleed", "rich black", answered from a table, instantly. A designer playbook ships in her knowledge so the routes she takes for a poster, a resize or a print hand-off are the ones a studio would.

“Export the active artboard as a 2× PNG and send it to the client on WhatsApp.”

Runs: needs-attention, active, history, with the transcript of each. Sessions: every Claude Code conversation on the machine, grouped by project, with the exact question a blocked one is waiting on. Specs: the design documents, answerable by voice. Projects: one project's conversations, runs and build progress. Usage: the subscription's two windows as gauges.

Connections shows every tool she can reach and whether it is ready right now, computed from what is installed and what actually started, never a capability she cannot back.

My Sites. Register your own websites and products by name, a portfolio, a client's console, a shop, and she can open them on a surface and, when a site publishes a data endpoint, answer "how is it doing" with live numbers, always saying the window they cover.

“Which sessions are waiting on me?”