Chat & your first edit

The chat panel is where you ask for edits. It is open on the right whenever a project is open, titled 28. Once the project has a transcript the header shows a “Transcript loaded” badge, which means the agent can cut, caption, and search by what is said.

The chat panel

• Type in the box labelled “Ask AI to edit your video…”. Enter sends; Shift+Enter adds a line.
• The round button sends. While the agent is working it turns into a square Stop button. Stopping ends the turn; edits already applied stay, and you can revert them (below).
• Header buttons: Chat history (saved chats for this project, with delete), New chat, and Close chat. Selecting an element on the timeline also closes the chat; the Chat button reopens it.
• An empty chat offers starter chips: “Add captions to my video”, “Cut”, “Add flying money when I mention earnings”, “Put a lower third with my name”.
• Long tasks show a live checklist that ticks off as the agent works, and a “Jump to latest” button if you scroll up.

Ask, Agent, Plan, Generate

The dropdown beside the input picks the mode. It is remembered per project.

When the agent asks first

• Proposed plan. In Plan mode the agent posts a numbered list with Approve and Cancel. After approval the same turn carries out the list. To change the plan, reply in the chat instead of approving.
• Cut preview. When the plan is a cut, you get a card reading “N regions · ~Xs to remove” with Apply cuts and Discard, a gear for the cut detection settings, and “Preview result on play (skip cut regions)” so you can hear the result before committing. The regions are also drawn in red on the timeline and can be dragged.
• Questions. If something is unclear, a question card appears with quick-reply buttons and a free-text field.
• Motion graphics. These are the exception to Plan mode: the agent describes the graphic it intends to build in prose and shows a single Approve button. Reply with changes to get a revised proposal. See Motion graphics.
• Story edits. Asking for a story edit from raw footage produces a plan card with a hook, sections, and durations, then Approve & build or Discard. If the project has no video type yet, you are asked “What kind of video is this?” first.

Reference images

Attach images with the Image button (PNG, JPEG, WebP, GIF), by pasting, or by dragging an image from the media panel into the chat. Each image gets a stable name in the conversation, shown on the bubble as image1.png, image2.png, and so on, so a later message can say “put image2 in the lower third”. Motion graphics can embed them; the rest of the agent uses them as reference.

Reverting a turn

Under every finished turn that changed the timeline there is Revert this turn's edits. One click undoes every change that turn made, including edits from an approved plan, and leaves anything you did by hand afterwards alone. The same row has Copy answer, Retry this message, and a usage popover showing the credits or API calls the turn consumed.

The AI tab on a selection

Select a clip, caption, or graphic on the timeline and open the AI tab in the properties panel. This is a chat scoped to that selection: “Zoom in slowly by 10% with smooth easing” on a video, “Duck it to -18dB” on audio. It has its own thread and per-message revert, and does not take images.

What AI can do

Cutting

Remove filler words, retakes, self-corrections, repeated content, off-topic speech, and silences; condense to a target length; trim or split a clip at a spoken word; reorder sections by what is said. All of this needs a transcript, which the Captions tab or the agent creates.

Captions

Generate captions, switch template, change grouping, restyle globally or per word, fix text, add emoji, move or scale the caption layer. See Captions & transcription.

Motion graphics and text

One designed graphic per request (a lower third, a stat callout, a quote) unless you ask for more, or plain text at a named position.

Media

Search your own footage by meaning once it is indexed; find clips, memes, GIFs, and stickers in the KLIPY library; find music and sound effects on Freesound; suggest stock and licensed web footage as cards you confirm before anything is imported. Meme, music, and SFX search are included with Plus and above.

Fine-tuning

Keyframes and zooms, effects, masks, transitions, snapping cuts to detected beats, smoothing jump cuts, reframing footage to fill the canvas, tracking an object so another element follows it, separating a person from the background, and ducking music under speech.

Export

From the in-app chat, exporting is done on the Export page. AI apps can start an export themselves; the file lands in ~/Movies/Agent28.

Chatting from ChatGPT or Claude

When editing from an AI app rather than the in-app panel, pasting this once at the start of the chat keeps the assistant on the tools:

Help me edit video in Agent28. Keep the Agent28 desktop app open while we work.

1. Confirm Agent28 is linked and a project is open (or help me open/create one).
2. When I ask for an edit, use Agent28 tools — don't invent timeline state. Prefer: cut fillers/silences, captions, motion graphics / lower thirds, import media, and export.
3. After each change, briefly confirm what changed on the timeline and how I can undo it from the chat panel.
4. If tools fail because Agent28 isn't running, tell me to open the app and try again.

Cut the ums and silences out of my Agent28 project.
Add captions to my Agent28 project.
Put a lower third with my name when I introduce myself.
Export my Agent28 project as a 1080p video.

Tip: Changes made from an AI app show up on the Agent28 timeline as they happen and can be reverted from Agent28's own chat panel, the same as in-app turns.