Chatting with Jarvis, Docs

Chat is the fastest way to work with Jarvis. Describe what you need in plain language, and Jarvis uses its tools, your Brain and your connected apps to get it done. For quick, exploratory or one-off work, chat is where you'll live. For work you want tracked and reviewed, hand it over as a card in Work.

You can also talk to Jarvis or message it in Slack. See Use Jarvis.

Starting a chat

Choose New task in the sidebar, or press ⌘N. Under the box, Jarvis suggests starters for your workspace and the time of day. Your conversations stay in the sidebar, so you can come back to any of them.

Voice, your Jarvis chat in the app and your Slack direct messages share one conversation, so Jarvis knows what you asked earlier. Jarvis also keeps notes across conversations in Brain → Memory.

Writing effective prompts

Jarvis works best when you're clear about the outcome.

Describe the outcome, not the step.

Instead of "help me with sales"
Try "Research Acme Corp and draft a cold outreach email to their VP of Engineering, focused on how we help dev teams ship faster."

Give it context. Name the details that matter (timeline, audience, tone, constraints) and attach files when Jarvis needs to see something.

Let it run, then steer. Jarvis chains many steps on its own, so you can hand it a whole job ("research, draft, and schedule the follow-up") and correct course from its output.

`/`-Skills: run a playbook

Type /, or choose Skills, to run a Skill by name: a saved playbook that tells Jarvis how to do a job your way. You can also just describe the task; if an installed skill fits, Jarvis uses it on its own.

Attaching files

Choose the attachment button, or drag files in:

Documents (PDF, Word, TXT, Markdown) and spreadsheets (CSV, TSV, XLSX), up to 10 MB each
Images (JPG, PNG, GIF, WebP), up to 5 MB each

Attachments are for one-off inputs, like a contract you just received or a screenshot. To keep something for later, add it to the Brain.

Picking a model

Each chat has its own model menu. Smart, the default, picks Opus, Sonnet or Haiku for each request. You can also pick a model yourself, and on Agently Pro you can run chats in the Mac app on your own Claude or ChatGPT plan.

Watching Jarvis work

Steps. Each reply can show the steps Jarvis took (searching the Brain, reading a page, drafting an email) with how long each took.
Sources. Facts from your Brain show From your brain with the sources behind them.
Approvals. Before Jarvis sends anything to people, deletes or archives anything, or merges code, it shows an approval card in the chat. Nothing happens until you choose. See Approvals.

Tips

Feed the Brain. The more knowledge you add, the sharper the answers.
Be specific about format. Want a table, bullet points or a formal email? Say so.
Iterate. Review, give feedback, ask for a revision.
Start fresh when it drifts. If a thread has gone off track, start a new one.
Promote repeat work. If you'll run it again, make it a Routine or a Skill.

Meet Jarvis, how the agent plans and acts
Use Jarvis, voice, approvals and background requests
Skills, the playbooks you run with /
Brain, the knowledge behind good answers