Your first conversation

Open a project, type a question, press Enter. That's the whole ceremony. This page is about reading what comes back.

What a turn looks like#

A turn is one thing you said plus everything the agent did before you spoke again. Turns are separated by generous space, and they render in this order:

Your message, in a raised bubble. Any attachments sit above your text as chips.

A left gutter with one dot per step, if the agent did anything besides talk. A turn that's pure conversation gets no gutter at all — the answer just flows.

The agent's thinking, collapsed behind a single word. Click it to read. It's there on every model, Claude Code included.

Tool cards for each thing the agent did — see What the agent can do. Above each one, a short line in the agent's own words says why it's doing it.

The answer, as plain markdown. No bubble, no avatar, no name label; the agent's prose is simply the text of the page. Code blocks are syntax-highlighted with the language named in the corner, and carry small actions: Copy on every block, ▶ Run in terminal on a shell command, and Open in browser on a block that is just a URL. A destructive-looking command is typed into the terminal but not entered, so you read it before it runs.

While the answer is still empty, nothing is drawn — the indicator above the message box is your only sign that work is happening.

Scrolling, and the thing that makes it feel right#

When you send a message, the view moves so your message sits at the top of the screen and the answer streams into the space below it. History glides up rather than snapping.

The transcript follows the bottom of the stream as it grows — until you scroll up. Then it stops following and stays where you put it. New output never yanks the view out from under you.

Four things start it following again: scrolling back to the bottom yourself, clicking the ↓ button that appears when you've scrolled away, sending another message, or switching to a different conversation and back.

Finding your way around a long conversation#

The rail on the right edge is one tick per turn — faint until you move the mouse near it. Hover a tick to see that message's text; click to jump to it.

Note

Clicking a rail tick stops the view following the live stream, so an answer in progress will keep going without you. Click the ↓ button to catch up.

Note

Very long conversations unload turns you aren't looking at, to stay fast. They come back as you scroll to them — but your browser's own find-in-page won't see text in the turns that are currently unloaded.

Things you can do to a message#

Hover any message of yours and a small toolbar appears under its bubble, right-aligned.

Copy puts your message text on the clipboard.

Resend sends the same message again as a new turn — or queues it, if the agent is busy.

Replay folds the conversation back to that turn and opens a stepper — see Conversations and history.

Restore rolls your files back to how they were before that message. Click once to arm it — the tooltip reads Click again to restore — undoes this turn and everything after — and again within three seconds to confirm; it posts a notice saying how many files changed. The messages stay — only the workspace moves. A checkpoint of the current state is taken first, so the restore itself can be undone.

Note

Restore only appears when there's a checkpoint to go back to. A checkpoint is taken the first time you approve a change in a turn, on every model — so a turn where the agent only read and talked has nothing to restore.

Clone copies the conversation up to that point into a new chat, so you can take a different path without losing this one. It's offered on the most recent turn.

Remove deletes that turn — your message and the response — permanently. Click twice.

Reading a tool card#

Tool cards carry the filename the agent was working on. Click the filename and the file opens in a tab, at the relevant line when the tool had one. It's the fastest way to check the agent's work as it goes.

The size of the text#

Theme → Typography → Chat font size scales the whole conversation. It's separate from the editor's font size and from Ctrl/Cmd +/−, which zoom the rest of the interface.

What the agent is told that you don't see#

A few things ride along invisibly, none of them in your bubble and none counted as something you typed.

Every message carries the agent's name and the current time — the clock, which session this is for the project, and how long since you last spoke. With memory on, it also carries what the agent already knows that's relevant to what you just typed.

The first message of a conversation additionally carries a compact map of your project — a file list, two levels deep, capped at 300 entries — so the agent starts with a sense of the shape of things instead of spending its first tool calls discovering it. With memory on, it also carries the hand-off note the previous session left, and whatever the project's other agents wrote on the shared whiteboard.

→ Memory — what it recalls and what it carries over