Chat with the Assistant

The Assistant is a conversation with Claude, built into the inspector of every editor window. You ask in plain words, and it reads and edits that window’s project.

Menu: View ▸ Assistant
Keys: ⌃⌘A
Where: Inspector ▸ Assistant tab
Needs: Your own Anthropic API key. Without one, the tab says how to add it.

The Assistant tab

You might ask:

“trim the pauses out of the interview”
“put a title over the first five seconds”
“what is under the playhead?”

The Assistant uses the same tools an agent uses over MCP, less a few held back from it, so its edits are ordinary edits. You watch each change land, and each one is its own undo step. See What the Assistant can do.

It’s the fifth inspector tab, after Clip, Media, Document and Transcript. Before your first message, save your key.

Open the Assistant

Choose View ▸ Assistant (⌃⌘A). Frame Job shows the inspector if it was hidden, switches to the Assistant tab and puts the cursor in the message field.

It isn’t a toggle. Pressed while the tab is already showing, it only puts the cursor back in the field.

The command stays enabled with no key, so the tab can say what’s needed. There’s no message field to put the cursor in until then.

The tab’s layout

From top to bottom:

Part	What it shows
Context	How full the conversation is, with Clear (↺) at its right end. See Long conversations.
Spend	What this conversation has cost so far, as Frame Job estimates it. See What it costs.
The transcript	Your messages, the replies and each tool call. See Read the conversation.
The ask-first row	Shows only while the Assistant is waiting for your permission. See What it won’t do and what it asks first.
The message field and Send	Where you type.

The tab stays put

The Assistant’s own edits change the selection without moving you off the tab, so you can read along as it works.

Clicking a clip in the timeline or on the canvas still switches to the Clip tab. A marquee on the timeline doesn’t.

Send a message

Click in the message field, which reads “Ask about or edit this project…”, and type what you want.
Press ↩ or click Send.

The field empties once the message is taken, and the reply arrives in the transcript. Everything the Assistant does between your message and the end of its reply is called a turn.

The field starts one line high, grows to five lines as you type, and scrolls past that. Spaces and blank lines at either end are dropped, and a message of nothing else does nothing.

You can type your next message while the Assistant works, but you can’t send it until the turn ends.

Message field keys

Key	In the message field
`↩` or keypad Enter	Sends the message and gives the keys back to the editor, so `Space` plays while the Assistant works. While it’s working or waiting for your answer, this beeps instead and the field keeps the keys.
`⇧↩`, `⌘↩`, `⌥↩` or `⌃↩`	Starts a new line
`Tab` or `⇧Tab`	Gives the keys back to the editor and never types a tab. With macOS Full Keyboard Access on, moves to the next or previous control.
`⌘Z` and `⇧⌘Z`	Undo and redo your typing

Clicking Send gives the keys back the same way ↩ does, and so does clicking anywhere else in the window below its title bar. See Keys and typing.

Your typing never lands on the project’s undo stack, and sending a message drops the field’s own undo history. While the keys are in the field, ⌘Z only undoes typing. Outside the field, it takes back the Assistant’s last edit.

↩ never answers a question and never stops a turn. While an input method is composing, ↩ belongs to it.

If a message cannot be sent

If the conversation can’t start, the reason appears in red in the transcript and your text stays in the field. Your message isn’t added to the transcript, so you can send it again once the cause is fixed.

The usual causes are a missing key and a helper that won’t start.

Send becomes Stop

The button reads Send, with the tooltip “Send the message”. While a turn is running or waiting on you, it reads Stop, with the tooltip “Stop the turn”.

Stop is a click. It has no shortcut. See Stop a turn.

What each message carries

Each message starts with a short block describing your window at the moment you sent it, so “this clip” and “the selected titles” mean what you’d expect. The block holds:

the project’s name, and its path, or a note that it’s never been saved
the nest open in the window, if there is one, or that the timeline is open
the playhead’s position
the ids of the selected clips

You don’t see the block in the transcript, but it goes to Anthropic with your message. See Assistant privacy.

It isn’t updated during the turn. The Assistant is told to read the project again after its own edits.

One conversation per window

Each editor window has its own conversation, about that window’s project. Closing the window ends the conversation and declines any question waiting for your answer.

Conversations aren’t saved with the project, so reopening it starts with an empty transcript.

The Assistant can’t close its own window. You do that with File ▸ Close (⌘W).

Clear the conversation

Click Clear (↺, “Clear the conversation”) at the right end of Context. It ends the conversation and empties the transcript, and your next message starts a fresh one.

Spend goes back to “Nothing yet”.
The context meter keeps the model’s window size, so after a conversation it reads “0 of … tokens · 0%”, with that size in place of the dots, until the next response.
Clear asks for no confirmation. It ends a running turn and declines a waiting question.

The project is untouched. Edits already made stay, each still its own undo step.