The Agent (Skills Chat)

The Agent is a chat assistant inside the editor that you specialize by loading skills — each skill teaches it a workflow (designing a car, building a storyboard, writing a video prompt) and it can generate images and videos and act on your canvas.

How the Agent works

The chat is the agent. One conversation = one chat. You load skills into a chat to specialize what it knows; the same generic assistant runs every conversation, and loaded skills extend it with specialist knowledge, optional workflow state, and image/video generation. You can load several skills into one chat and unload them at any time.

Open / collapse the Agent

  • Open it with the robot button at the bottom of the editor (toggles Open agent / Close agent).
  • Inside the panel, the header button toggles Show chat / Hide chat (also labelled Collapse chat) to fold the panel down to just the composer and back.

Start a new chat

  • In the panel header, the + button (New chat or reopen recent) opens a popover with New chat at the top and your recent (closed) chats below.
  • On the landing screen, Create new (Start a fresh chat without a skill loaded) starts a blank chat, or pick a skill bubble under Or start with a skill to start a chat with that skill pre-loaded (its intro message appears first).

Load / unload skills

A skill specializes the chat. Three ways to load one:

  • Slash picker: type / in the composer to open the skill list and load a skill into the current chat. Already-loaded skills show a Loaded pill.
  • Suggestion bubbles: on the empty/landing screen, click a skill bubble (under Or start with a skill) — this starts or seeds a chat with that skill and posts its intro message.
  • Auto-load: the agent can quietly load a relevant skill itself mid-conversation when your request matches one.

Loaded skills appear as chips above the textarea. Click the × on a chip to unload that skill. A chat can hold multiple skills at once.

Send a prompt

Type in the composer and press the Send button (or Enter). A voice-input button is available for dictation.

Attach image / video references

  • Click Add image (+ / New image) in the composer to pick a reference from your library via the image dialog.
  • Type @ in the composer to mention an image from your library inline (e.g. @image1).
  • Attached references appear as a thumbnail strip above the textarea; each tile has a Remove control, and clicking a tile inserts its @-token into the prompt. Both image and video references are supported.

References you attach are carried into generations the skill runs (e.g. locking a character, a vehicle, or a style).

Thumbs feedback on generated images

Generated images show Thumbs up / Thumbs down buttons. For iterative skills (e.g. Car Designer), your votes steer the next batch: thumbed-up images are pinned and the next generation sweeps around them. Click an active vote again to remove it. Use View full size to open an image in the lightbox.

Confirm a generation

When the agent is about to run a video generation, it shows a Yes button (Confirm and proceed) with an estimated credit cost (≈ N) and, on hover, a cost breakdown. Video generations only fire after you confirm.

Image generations do not go through that gate. Asking for one or two images runs them in the same turn, with no button and no cost shown beforehand — the agent is instructed to skip the confirmation at that size and ask first only for larger batches. The credits are reserved when the generation is submitted, and the spend appears in your credit balance afterwards.

Choose from multiple next actions

When the agent offers several next steps, it renders a row of choice buttons under the message. Each shows its label and, when costs differ, a per-action credit estimate. Click one to proceed with that action.

Apply agent canvas actions

When the agent proposes canvas/project changes, an action button appears (e.g. arranging selected nodes) with a preview of what it will do. Click the button to apply the actions to your canvas. Tip: select a few nodes and tell the agent something like "arrange these in a grid".

Add a generated image to the canvas

Drag any generated (or attached) image tile from the chat onto the canvas to place it. A tile already placed shows an On canvas chip. Images live in the chat's Gallery view too — toggle between Chat and Gallery with the view button in the header (layout button switches between single/large and grid).

Continue in a fresh conversation

When a chat fills up its context, a Continue in fresh chat button appears (and a notice that the context limit was reached). It spins up a new chat that carries over your loaded skills, workflow state, and pinned references, plus a short handoff summary — so you keep working without losing the thread.

Manage chats (tabs & history)

  • Open chats show as tabs; each tab has Rename chat and Close tab controls.
  • The + popover (New chat or reopen recent) lists recent closed chats — click to reopen, or use Delete to remove one permanently.
  • Use Open an existing chat from the landing screen to jump back into a past conversation.

Available skills

Load one (or several) to specialize the chat:

  • Car Designer — iterative car ideation loop: ships 4 image variations per turn; thumb them up/down and the next batch sweeps around your picks. Supports a design-reference image (pull in the look of any object) and a tuned car LoRA. For designing a vehicle through visual iteration.
  • Character Creator — design a character (humans, anime, comic, anthropomorphic animals, fantasy, robots). Walks through type → identity → outfit, then ships a white-seamless base reference plus a 6-panel character sheet for consistency.
  • Storyboard — turn an idea into a storyboard / shot list / pre-vis for a film, ad, or music video; breaks a concept into beats and visual shots.
  • Cinematic Prompt — compose a video prompt for general cinema. Locks characters/wardrobe/environment by visual description across five cinema modes (Narrative, Studio, Action, Performance, Atmospheric).
  • Cinematic Prompt — Seedance — the same composer tuned for Seedance 2.0's deep reference stack (multiple image, video, and audio references in one generation).
  • Cinematic Environment — image-prompt composer for environment plates: single photoreal locations with no characters in frame (architecture, set dressing, atmosphere, light). Pairs with character work and video skills.
  • Car Commercial — automotive video layer on top of the cinematic composer: launches, dealership content, hero turntables, drive footage, track/rally clips. Locks the vehicle by visible properties and can run reference-to-video against the chat's references.
  • Car Interior Designer — iterative car-interior ideation loop (dashboard, IP, seats, center console, headliner, door cards, trim): ships 4 cabin variations per turn, thumb-driven, with material/palette design-reference support and a tuned interior LoRA. The car-exterior loop's sibling — same verbs, different vocabulary.

Related