Docs menu

The Ask AI assistant

The Ask AI panel: one chat for the whole film that finds problems and fixes them, with approval cards, a price on every paid step, auto-approve and a spending cap.

Reference7 min read

On this page

While you work on a film you will often want a helper that can find a problem and fix it for you. The Ask AI panel is that helper: one continuous chat for the whole film, with the shot you have selected already in context. You describe what you want, it inspects the shot and its neighbours (and the rest of the timeline when it needs to), tells you what it found and applies the fix. Every step that costs money waits for your approval, with its price on it.

Opening the assistant panel

When something in the film is wrong, you reach the assistant, Fixer AI, in three ways: the Ask AI icon on the side panel's rail (the fourth icon, under Characters / Locations / Assets), the Ask AI pill above the shot editor, or the Fix it button in a shot's quality warning. It is one continuous conversation about the whole film; you type what you want into the box at the bottom and follow the answer in the transcript. If you switch to another tab, a running job keeps working.

  • Ask AI: opens the conversation inside the side panel. While Fixer works and you are on another tab, the rail icon turns into a spinner; replies you haven't seen pile up as a "+n" badge.
  • Fixer AI (the panel title): there is one conversation per project, so the panel carries the product's name. Beside it: the ⚡ auto-approve menu and the New conversation icon.
  • The empty state: a fresh conversation shows a plain question, "What should we fix?", one line on scope ("One shot or the whole film. The shot you select is in context automatically.") and three starters: "Two people's mouths move at once. Fix it", "Darken the lead's outfit in scenes 12–18" and "Render every shot that has no video yet". Clicking a starter only drops its text into the message box so you can edit it; it never sends by itself.
  • The message box ("Ask anything about the film…"): one soft box: the text on top, and inside it a row with the context switch on the left and Send on the right. Up to 2000 characters; a live counter appears near the limit and longer pastes are trimmed. Enter sends, Shift+Enter adds a line break; sending is disabled while an answer is streaming.
  • Thinking…: shown while the answer is being prepared; a faint "AI working" wave runs behind the transcript.
  • Read steps (list shots, shot detail, search, cast and locations, production status, cost estimate, balance, reference board, prompt suggestion): the assistant's information-gathering shows as faint rows; they need no approval and are free. Click a finished row to expand its result.
  • Fix it: in the quality-warning dialog of a shot (from the card's yellow badge or its right-click Findings and notes), hands the failed checks to Fixer as they are: the panel opens on that shot with the findings as the first message. If you approve, the shot is regenerated, and the warning clears itself after a clean render.
A fresh conversation: the ⚡ auto-approve menu and the new-conversation icon in the header, the question and three starters, and the composer with its Whole film | Shot 1 switch and send button.

One conversation, and the context switch

There is ONE conversation per project and it runs continuously: fix a shot's mouth sync, then say "now tidy up the ending", and the assistant remembers both, because they happened in the same conversation. The whole film is always in context: the assistant sees the shot map, the cast, the locations and the production status, and reads individual shots itself when it needs them.

  • The context switch: inside the message box, bottom left: Whole film | Shot N. The second half appears whenever a shot is selected in the timeline and carries its number. With it on (the default once you select a shot), that shot's full detail and its frames go to the assistant, so "this shot" or "the man here" has a referent. Click Whole film to leave the shot out of the next message without deselecting it in the timeline; click the shot half to put it back. Selecting a different card moves the switch to that shot. With no shot selected, only Whole film shows.
  • New conversation: the pencil icon in the panel header. Closes the running conversation and opens a blank one; disabled when it is already blank. You will rarely want it: the continuing conversation is exactly what stops the assistant from proposing a fix it already watched fail.

The conversation is stored in your browser. Starting a new one discards the old transcript; edits the assistant made with your approval can still be reverted with the editor's normal undo.

Approval cards and the price tag

Fixer never spends money or changes content without telling you. Every tool call that spends or modifies stops as an approval card in the transcript and waits for your decision.

  • Approve / Discard: the card shows the operation's name in plain language (e.g. "render the video", "edit the scene"), the AI's reasoning and a price tag. Approve applies it, Discard rejects it (the row stays, marked as rejected). Approved edits are also pushed onto the editor's undo stack, so ⌘Z can revert them.
  • Price tag: the amount comes from the very same calculation the server will charge your wallet with, written as dollars + token count together. Free operations say "Free"; open-ended jobs that cannot be priced say the amount depends on how far the work goes. The prefix "at least" means the amount can rise; "approx." means the engine is set to Auto and the price is based on the table default. Multipliers that raise the amount are listed by name: "two named characters: a second pass re-locks every face" (×2), "covered hair: a second pass fixes the headscarf" (×2), "includes the thumbnail and face-lock every model page produces".

Free operations such as text and cast patches show as faint rows instead of cards; only generation tools (image and video renders, audio generation, cover art, frame edits, mouth sync and the like) spend money.

Auto-approve and the spending cap

If you don't want to approve a stream of small fixes one by one, switch the conversation to auto-approve. Fixer then applies its fixes without asking until the conversation's spending cap is reached. Destructive operations are still always asked.

  • Auto-approve off / Auto-approve on: toggled from the ⚡ menu in the panel header. "Fixer applies its fixes without asking, until this conversation reaches its spend cap. Deletions and starting the pipeline always ask." Off by default; while on, the ⚡ button shows the remaining budget in dollars.
  • Spend cap for this conversation: inside the ⚡ menu: 2,000 / 5,000 / 10,000 / 25,000 tokens, each with its dollar value; the default is 5,000 tokens ($1.56). Under it a ledger line reads "Spent: … / …"; when the cap is hit, "Cap reached. Fixer is asking for approval again."
  • Approved automatically: a spend nobody clicked shows no card; the running or finished row carries ⚡ and the amount instead. The transcript doubles as the conversation's spending record.

The safety rules live on the server: the cap is clamped to 25,000 tokens whatever is requested; operations that can't be priced are never auto-approved; deleting a shot, a character/location/asset or an audio clip, reordering the timeline, shortening the film and starting production always ask; and at most 12 operations are auto-approved in one turn.

Costs and limits

In the assistant, only generation tools cost money.

  • In Fixer, only generation tools (image and video renders, audio generation, cover art, frame edits, mouth sync and the like) spend money. Each approval card shows the price in dollars and tokens, identical to what the server will deduct. Read steps and text or cast patches are free and pass as faint rows.
  • With auto-approve on, the price shows as ⚡ + amount beside the running row. The cap is enforced on the server and clamped to 25,000 tokens; deletions, reordering, shortening the film and starting production are still asked every time.
  • On team projects every generation is charged to the project owner's balance; only the Editor role can start production. If the owner's balance runs out, a member sees a notice that the owner needs to top up.
  • Fixer's limits: 2000 characters per message (the server trims too) and 20 messages per minute per user. Beyond that you see "Too many messages. Please wait a moment and try again." Fixer stays on topic and declines attempts to extract its instructions.
  • Starting a new conversation can't be undone, but the assistant's approved edits can still be reverted with the editor's undo.