Skip to content

Assistant

The Assistant is a chat panel that knows your project and can act inside it: set up channels, kick off generations, edit images by instruction. Toggle it from the header and it docks on the left in any mode.

Note

This is different from Assistant Editor, the sub-mode inside Video mode that turns a Redub transcript into a paper cut. Same word, unrelated features.

It needs a provider key for whichever brain you pick: OpenRouter, Google (Gemini), OpenAI, or Anthropic (Claude). OpenRouter can route through your Fal key via the Use Fal for OpenRouter setting. See Connect providers.

Sessions

Conversations are organized in a session tree: make folders, rename sessions, drag to organize. Sessions are scoped to the project, or to a specific Kitlig Asset, so the goblin's art-direction thread stays with the goblin.

At the top, the Agent select switches personas: the "Default assistant" or any preset you've made (see below).

Chat

Type into "Message the assistant..." and press Enter. Replies render as markdown, and you can attach an image to a message. When the Assistant does something in the app, its tool steps show as check lines in the thread, so its actions are visible rather than hidden.

Hovering a message offers Copy, Rerun, Edit, Fork (branch the conversation from that point), and "Save as text asset". Select any text in a reply and "Save selection" captures just that portion. Saved text goes through a small dialog (Name plus Folder, defaulting to text/generated) and becomes a regular text asset.

What it can actually do

The honest capability list:

  • List and create Kitlig Assets
  • Set a channel's prompt and pipeline
  • Start generating a variation
  • List and create agent presets
  • Edit an image by instruction (see below)
  • Answer a job that paused on a question, or relay the question to you (see Jobs)
  • Describe and tag images and video (a vision model writes what each file shows, with timestamped moments for video), and find footage by those descriptions
  • Organise the project by those descriptions: bins, consistent tags, Kitlig Assets, and folders on disk when you ask for them. The built-in Asset Organizer preset is made for this; it works through every matching asset, not a sample, and reports each group with its count
  • Read text files in the project (transcripts, subtitles, JSON side-cars)
  • Edit the Video timeline (see below), including turning an audio lane's speech into a subtitle lane

Editing the timeline by conversation

With a sequence open in Video ▸ Edit, the Assistant can read and change it: lay footage, stills and titles on lanes, trim, split, move and delete clips, close gaps, add markers, set the frame size, save and switch versions, create bins, and render an export. It works on the open sequence unless you name another one. The built-in Video Edit Assistant preset carries the editor's instincts; the plain Assistant has the same tools.

Things worth knowing:

  • Ask in editing terms: "put a lower third for Kam at 3 s on V2", "cut the first clip where she says storm", "move everything in the Food bin after the interview", "export at 1080×1920".
  • While it works you see what it is doing: each finished tool call appears under the message as it happens, with the current step ("Running analyze_assets…", a render's own progress line) and the time the turn has been running. The same steps stay under the reply afterwards.
  • Every change it makes is an undo step: Ctrl+Z in the Edit view rolls an Assistant edit back like one of yours.
  • At the end of each turn its edits are saved as a minor version (1.1, 1.2, …) with a description of what it did; when it hands back a finished cut it saves a major version. The Versions menu on the toolbar shows them all, with an "agent" chip.
  • Bins are how you point it at footage: put the interview takes in an "Interview heads" bin and the cutaways in "Food", then say "edit everything in Food around the interview heads".
  • It cuts by what the footage shows when the assets carry descriptions (see Describing assets with AI): "use the shot where the plate lands" works because the moments are written on the clip. Ask it to "describe the Food bin" and it runs the analysis itself, one paid call per clip, asking first when there are more than a few.
  • "Make subtitles for A1" turns that lane's speech into caption clips on a Subtitles lane, transcribing locally where no transcript exists.
  • An export it starts is a normal Export Media job; the file appears in the project like any other render.
  • For music-driven reels there is a dedicated persona and workflow: Reel editing with the Assistant.

From the Files panel or a variation, right-click and choose Discuss in Chat to bring an image into the conversation.

Type @ in the message box to point at something by name: Kitlig Assets, files in the project, bins and sequences appear in a list that narrows as you type; arrow keys and Enter (or a click) drop the name into your message. The Assistant receives the exact ids behind those names, so "trim @interview_take2.mp4 to the first sentence and put it in @Food" works without it having to search. A mentioned image or text file also comes along as an attachment.

Image edits by instruction

Ask for an edit in words ("remove the background figure", "make the jacket red") and the Assistant runs an image edit for you. Quality has three tiers (Cheap, Normal, High), each mapped to a pipeline under Settings > Tool Models in the "Image Edit" row. Cheap for drafts, High when it's going in front of a client.

Plan

The Plan tab turns a one-line brief into a per-channel starting point:

  1. Describe the thing in Goal ("a swamp shrine environment kit, hand-painted style").
  2. Click Generate Plan. You get editable prompts per channel.
  3. Edit them like any brief, then either Save Plan as Text Asset or apply it: pick a new or existing Kitlig Asset and click Apply & Open.

The plan fills the channels' prompts; nothing generates until you say so. It's scaffolding, not autopilot.

Agent presets

Presets are saved personas: a lore-keeper that answers in your game's voice, a strict art director, a naming generator. Open the Agent / Model Presets dialog to manage them.

Each preset has a Name, Provider, Model, System Prompt (the standing instructions that define its personality and rules), Description, Category, and Tags. Presets import and export for sharing (see File formats), and the built-in ones are read-only. There's also a built-in "Agent Creator" persona under Settings > Agents, with its own provider dropdown, whose job is helping you write new presets.

The other built-in is "Lyric Smith", a songwriter. Give it a mood, a title, or a situation and it writes the lyrics, then lists the matching settings for Audio > Music > Generate Song (genres, era, country of origin, length, BPM, vocal accents). Ask it to make the song ("generate it", "let's hear it") and it starts the generation itself. The song lands in your project as a new audio asset with its lyrics beside it as a text asset. Each generation costs about $0.10, so it only generates when you ask, and offers first when you only asked for lyrics.