The AI Assistant
What the AI Assistant can do, typing and voice typing, and the skills it reaches for.
Ordinary Animator includes an AI Assistant that can answer questions about your production, navigate to any entity, retrieve data, and help with production tasks. You ask it by typing, or by holding the microphone and speaking your message instead of typing it. There is just one Assistant, and you never have to pick who to talk to -- behind the scenes it reaches for whichever of its skills your request calls for (see Skills below).
The right-hand column has three views, and a Chat | Docs | Info switch at the top flips between them in one tap. Info reminds you what the shot, scene, episode, character, location or prop on the page you are looking at is: a few plain lines such as what a shot is for, what happens, the camera, who is in it and where, its dialogue, length and visual style. It follows the page as you move, leaves out anything not yet written, and costs no AI. Chat is the Assistant. Docs is the written description of a field: what it is for, an example, and a link to this guide. The column remembers which view you last used. The Docs view only ever changes when you ask it to -- moving to another page never changes what it shows. Ask in any of these ways: type in the search box at the top (it looks through every field's name, card, description and example); press This page's fields to list the fields of the page you are on; pick another kind of item under Browse fields of; press the book icon beside a field on a properties page or in the plan; or press Explain in the guide bar. Fields are listed under the same card headings the properties page uses, and each one is a row you can open and close -- Expand all and Collapse all open or close the lot, so you can open a few and scroll through them. The book icon and Explain open the column, flip it to Docs, and scroll to that field with it open.
One of the most useful everyday moments is field review: when you have written something in a property and want a second opinion, the sparkle next to the field asks the Assistant to Analyze it. It reads your current value, quotes it back, judges it against the field's description and your other properties, and offers a better version.
The sparkles and Suggest and Review with AI buttons dotted around the app are shortcuts to conversations like that. Each one writes a single plain question into the chat as though you had typed it -- "Analyze my 'Character role (story function)'.", "What should I make next here?" -- so your transcript reads back as a conversation you had, and you can edit the question or ask a follow-up in your own words. You never have to tell the Assistant which page you are on or which thing you mean: it already knows.
What the Assistant can do
The Assistant has a set of tools it can call:
- Navigation -- open any page (project, episode, scene, shot, character, location, prop) by name or ID, including a particular page of one: "open episode 2's Shot List", "show me Venn's Voice". Ask for a page that entity does not have and it says so, and names the ones it does, rather than opening something else
- Entity queries -- list all characters, find a location by description, retrieve a shot's details
- Data retrieval -- read screenplay text, look up voice settings, check what media a shot has
- UI commands -- trigger actions on the current page
- Workflow diagnosis -- look at a workflow of yours that will not run, including the one behind your last failed job, and say in plain terms what is wrong with it
- Making a workflow -- when a model in the catalogue has no workflow of its own, make one for it and add it to your project (see Making a workflow for a model that has none below)
- Generating media -- set an operation up in a gallery's Generate tab and run it, then tell you how it went
- Moving media between galleries -- link or copy a picture from one gallery into another, the same two transfers the gallery's Organize panel offers: "link Andara's sad close-up into this shot's gallery", "copy it to George's gallery". Name the item and the destination; a link borrows the file and a copy gives the destination its own entry (see Linking media between galleries)
- Choosing a shot's storyboard and clip -- set a picture from a shot's gallery as its storyboard, or a video as its rough clip, exactly as the card menu's Set as storyboard and Set as rough clip do: "set the newest image as the storyboard", "set that clip as the rough clip". The storyboard has to be an image and the clip a video from that shot's own gallery; ask it to clear either and it does
The Assistant works from your saved project data. It does not learn from conversations on its own; the only things it carries from one conversation to the next are the ones you ask it to remember.
Running operations for you
Ask for something a gallery can make and the Assistant does the setting up: it picks the operation, fills in the inputs it can work out, names the ones it cannot, and asks whether to run it -- and whether to run it on Comfy Cloud or your own ComfyUI, because that spends either credits or your own machine's time and is not its decision to make.
It picks the operation; your visual style recommends the model. Ask for a keyframe on a shot and the Assistant sets up keyframe -- which workflow actually runs it is the shot's look answering, exactly as it would if you had set the operation up yourself. So a shot in a dream-sequence scene gets the dream look's recommended model and the words that look adds to the prompt, with nothing to explain and no question to answer. The run row names the look it used. If you name a workflow yourself -- "use SDXL for this one" -- that wins, and the run row says it was chosen for that run rather than naming a look. And where the shot has no style, or its style recommends nothing for the operation, the Assistant says so and asks you to pick rather than quietly choosing for you. It never creates a style unasked: when it plans a scene's shots it only chooses among the styles you have defined, with a reason.
All of that happens in the Recipe panel of the gallery the operation is for, so you see the operation and its filled-in inputs before anything runs, and you can change a value yourself before saying yes. (The prompt box on that panel has its own Write prompt button, which writes the prompt the same way -- see Write prompt.) Ask from somewhere else -- the scene page, a different shot -- and the app follows the Assistant to that gallery and the run starts in front of you; going there is part of the job, not something you have to do first. The browser does the work and the Assistant only asks for it, so a gallery page does have to be open -- but it does not matter which tab of that page you are on.
A gallery's Recipe panel shows its recipe as a list of steps, and the Assistant adds to that list. Ask for something else after a result and it appears as another step under the first, which you can open, change and run yourself.
So something that takes several operations in order -- a still, then a clip from it, then a cleanup pass -- is a recipe of steps: the Assistant sets them up with each result fed into the next and runs them (the same as pressing Run all), or runs one, waits for the result and sets the next one up with it. Either way you can see every step, and the Graph view afterwards shows the whole run as arrows from one item to the next. See The Recipe panel. Changing the sequence is another sentence rather than a job for you: "same again but skip the upscale".
It tells you where it has got to as it goes, because a video operation can take twenty minutes, and it reports what actually happened rather than assuming. Stop ends a run after the operation in flight; a render already sent to ComfyUI cannot be recalled and finishes on its own. To stop the render itself, ask: "stop the current Comfy job" cancels it on your local ComfyUI or Comfy Cloud, without taking you off the page you are on, and the job's row in the Comfy Queue reads Cancelled.
A whole scene at once
"Storyboard every shot in this scene." "Make videos for all the shots." One sentence covers a whole scene: the Assistant works through its shots in shot-list order, and for each one sets the operation up on that shot, moves the app to that shot's gallery and runs it there, then goes on to the next. You watch it happen -- one run at a time, in front of you, with a card counting "Shot 3 of 7: 3.2 Eyes power on" so you can see how far it has got. Each shot's own look picks the model as usual. Each shot's prompt is written the way the Write prompt button writes it -- from that shot's camera, action and location directions, its line and the model's own guidance -- and the Assistant tells you what it wrote. The Assistant does not compose prompts itself; the only words it passes on are ones you gave it ("make shot 2 slower, she is unsure").
A shot that already has one is left alone. A shot whose storyboard still is already chosen, or whose gallery already holds a picture of the kind you asked for, is skipped and reported as skipped -- so asking again after adding two shots does the two new ones and nothing else. If you do want them all made again, say so ("redo all the shots") and it will, which spends credits on pictures you already have.
A shot that fails does not stop the scene. It carries on to the next one, and when it finishes it tells you which shots were made, which already had one and which failed, by shot number, so you know exactly what to look at. Stop ends the walk after the shot in flight and says which shots were never started.
Animating a scene works the same way: "make videos for every shot" starts each clip from that shot's own storyboard still -- the one you marked as the shot's storyboard, not whichever picture is newest -- and a shot with no still yet is reported rather than attempted.
A shot that has a line is animated differently, in the same single sentence: "make videos for every shot, dialogue or not" gives each talking shot two steps -- the character speaks the line from the shot's storyboard (which stays exactly as drawn), then the clip is given that character's own voice -- and animates each silent shot plainly. The first clip is as long as the line. The voice is the character's voice sample from the Character Voice page, and nothing else -- a TTS voice alone is not a sample. A character with no sample is reported by name, before anything is paid for, instead of being given someone else's voice: "Sally has no voice sample. Make one from the TTS voice, or upload one on the Character Voice page." The Assistant offers to make one (with a Make a sample suggestion) and only does so on your yes -- then it makes the sample the way the Make a voice sample button on the Character Voice page does (the same sentences and the character's TTS voice) and carries on with the talking shot you asked for. When two characters speak in turn in a shot, the steps are Talking shot, two voices instead: the clip speaks both lines, then each speaker's part gets their own voice sample with the ambience kept, cut at a split point the Assistant fills in as an estimate for you to correct; three or more speakers in one shot are reported rather than run. A talking shot's render that lost the storyboard's picture is never counted as the shot's rough clip; it is reported as a rejected render. A shot whose dialogue audio has not been rendered yet is animated silently, and the Assistant says so: press Render dialogue audio on the shot and ask again. A lipsync run you ask for by name ("do lipsync on all the shots") uses the clip-driven models instead: each clip is made from the shot's storyboard and the dialogue audio rendered for its speaker, and a shot with no rendered audio is skipped, with that reason. If the look's lipsync default is a workflow that has since left the catalogue, a shot with a rendered line takes the talking route above instead, and the report says which route each shot took; a shot with no rendered audio is reported, never animated plainly. To redo a shot's video, say "regenerate shot 1.2": that is the talking route, not a lipsync model.
Which recipe makes a shot's video is the show's choice, and the Assistant names it in the report ("Talking shot, by this show's choice: ..."). If more than one recipe fits a shot and nothing is chosen yet, it asks you once, which to use and for what, and remembers your yes; "use the two-step route for this shot" remembers it for that shot only. See Models, Workflows & Recipes.
Several browser tabs open at once
People keep the same project open in more than one browser tab. When you ask for a run, exactly one of them does it -- and it is chosen rather than whichever happens to answer first.
The tab that takes the job is one that is showing that gallery, can reach the backend the run needs right now, and is running the newest version of Ordinary Animator among the tabs you have open. Of those, the one you are looking at goes first. The others stand back, and step in only if the chosen one has gone away.
That matters most for your own ComfyUI, which you start and stop as you please. A tab left in the background can be holding a stale answer about whether it is running; it no longer gets to refuse a run on the strength of it. And if genuinely none of your tabs can reach the backend you asked for, you are told so straight away, naming which one is missing -- rather than waiting for a run that was never going to start.
A tab that has been open across an update is kept out of the way for the same reason: it is still running yesterday's version. It tells you a newer version is available and to reload, and up-to-date tabs handle runs in the meantime. It will reload itself only when that costs you nothing -- when it is in the background, you have nothing unsaved in it, nothing is running in it, and nothing is part way through. If you have typed something into it, it waits and leaves the choice to you.
Making a workflow for a model that has none
Most of the models in the catalogue have no ready-made workflow of their own. That used to mean you could not use them without building one yourself in ComfyUI. Now you can just ask.
When a model does the same job as one that already has a workflow -- another upscaler, another image model of the same family -- the Assistant makes the new workflow from that one. It copies the working version, points it at the new model's file, and gives it a name and description of its own. Everything else about it is untouched, which matters: the settings, the inputs you fill in, and the way the result is saved are all exactly what made the original work.
The new workflow is checked before it is kept -- the same checks any workflow of yours goes through -- and only then added to your project's workflow library, where it appears in the pickers like every other workflow. You do not have to accept anything.
That makes a comparison a single sentence. "Try all the local upscalers on this image" is the Assistant making a workflow for each upscaler that does not have one, then running them one after another on your picture and naming each result for the model that produced it, so you can look at them side by side and pick.
When a model is not like anything that already has a workflow, the Assistant builds one from scratch instead -- and if something it needs is genuinely missing, it says exactly what: a node that is not installed where you want to run it, or a model file nobody knows where to get. It never invents either.
Asking what to do next
Ask the Assistant "what should I do next?" and it answers from where you are standing. If the page you are on has open work, that is what it offers first -- the character, location or scene in front of you, not whichever one the plan happens to list first. It then names at most one alternative, so you get a choice rather than an order: finish this here, or start that instead. Characters, locations and props have no required order between them, and the Assistant will not invent one.
When the page you are on is finished, it says what is open elsewhere, so you are never told there is nothing to do while work remains. If the page is waiting on something earlier -- a scene with no shots yet, say -- it names the thing that has to happen first. It is reading the same open work the guide bar at the top of the page is showing you, so the two always agree.
Its answer comes from the episode's progress report, which is worked out from your plan, not improvised: where each phase stands, which steps must come first and why, and a short ranked list of what to do next. Steps whose outcome already exists are left out -- once your screenplay is written, the Assistant will not send you back to build the story outline it was written from. Open the same report yourself with the Progress report icon on the guide bar.
Answering with one press
When a reply ends by asking you something with a handful of short answers -- "shall I run that on the cloud or on your local ComfyUI?", "shall I run it?", "the first or the second one?" -- a row of small buttons appears under it, one per answer. Press one and it says that word for you: it arrives in the conversation as your message, exactly as if you had typed it, and the Assistant carries on from there. You can always ignore the buttons and type something else instead, including something that is not one of the choices.
The buttons only appear where a choice was actually offered, and only under the newest reply -- once you say anything, the old row goes, because an answer to a question from three turns back would land in a conversation that has moved on. A half-written message in the box is left alone when you press a button.
Showing the Assistant a picture
The Assistant can already show you a picture in its replies -- ask what is in a gallery and the items come back as thumbnails you can click. You can do the same thing back: drag an item out of any media gallery and drop it into the chat box.
A small thumbnail appears above the box to show what you have attached; the box itself keeps only your words. Refer to it the way you would in conversation: "what is wrong with the attached image?", or just "what is wrong with this?" when there is one. Drop two and they stay in the order you put them in, which is what lets you say "use the first one as the reference". Send the message and it appears in the transcript with the pictures under your words. The cross on a thumbnail takes it off again before you send.
The Assistant answers about that item: it knows which one it is, what kind of file it is, and can feed it straight into an operation. You never have to describe the picture or tell it which gallery to look in -- an item dragged from a character's gallery onto a shot's page is still that character's item, and the Assistant treats it as such.
So an attachment can be handed straight to an operation -- "remove the background from this" sets the operation up and runs it once you say where. You watch it happen in the gallery's Recipe panel: it switches to Remove Background (RMBG) with the item you dropped already in its Source Image, and the run starts there without your pressing anything. Because the Assistant works from the item rather than from a copy of the picture, the result remembers where it came from: the gallery's graph view draws an arrow from the item you dropped to the new one.
To take one off again, click the small cross on its thumbnail. You can attach up to eight items to one message.
Nothing is uploaded: this points at media that is already in your project. To bring a new file in, add it to a gallery first.
On a tablet, use the item's menu
Dragging needs a mouse. On a touch screen there is no drag to make, so every media item's "..." menu carries Add to chat -- in the grid, the list and the graph alike. It attaches exactly what a drop attaches: the same thumbnail above the box, in the order you added them, referred to the same way.
The item has to have somewhere to go, so the action appears only while the chat panel is open. Open the chat first, then add.
Dragging a picture out of a reply
It works in the other direction too. Every thumbnail the Assistant shows you can be dragged out of the chat panel, because each one is a real item in one of your galleries, not a picture of one.
Drop it on a gallery -- a Media page, or an experiment -- and it asks one question, Link or copy?
- Link shows the original. The two places share the same file, and the gallery makes clear the item is a link, so you can see where it really lives.
- Copy makes an entry of its own. Use this for something you want to keep after whatever produced it has gone -- an experiment result, for instance, since deleting the experiment would empty a link.
Some items can only be copied: nothing can be linked out of the project gallery, an episode gallery or an experiment, and the Link button says so when you hover it. Dropping an item back on the gallery it already lives in does nothing, and so does dropping one that is already there.
You can also drop a thumbnail straight onto an operation's image input in the Recipe panel. That asks nothing -- the box fills, exactly as it does when you drag a card out of the gallery beside it.
Or just select it
You do not have to drag. Click an item in a media gallery and the Assistant can see it -- in the grid, the list or the graph, and the one you have open full screen counts too. A quiet line above the chat box says so ("1 selected image will be included"), so it is never working from something you cannot see it is working from.
Say "remove the background of the selected image" and it acts on that one. Ctrl-click to add a second, shift-click to take a run of them, and ask "which images are selected?" to hear what it is holding. Click away or press Escape and the selection is gone, and so is what the Assistant knows about it; leave the gallery page entirely and the same thing happens.
Selecting and dragging mean the same thing to the Assistant, with one difference: if you have dropped something into the chat box, that is what "this image" means, and the selection waits until you ask for it by name -- "use the selected one as the reference". As with attachments, up to eight selected items travel with a message.
Asking it to actually look
Pointing at a picture is one thing; the Assistant looking at it is another, and it can now do that. Ask what an image shows, or ask it to compare two, and it sends each one to a vision model and answers from what came back.
Comparing is the case it is built for. Select two results and ask "compare these -- which looks more like George?", and you get a short verdict on each one, the differences that actually matter for the question you asked, and a recommendation. It works from the pictures, not from their names or the operation that made them -- so "which of these is sharper?" is a question about sharpness and not about filenames.
Each answer is kept on the item, so asking the same thing again is instant and costs nothing. Asking something narrower -- "is the collar open on either of them?" -- is a new question, so it looks again.
Two limits worth knowing. Images only: ask it to compare two video clips or two dialogue takes and it will tell you it cannot watch or listen yet, rather than guessing from the filename. And a look costs a fraction of a cent, which is why it looks at the items you name rather than at a whole gallery.
Comparing inside an experiment
An experiment already has a question written on it, so a comparison made inside one is answered against that question without your having to repeat it. Generate a couple of results, select them, and ask which one wins: the summary ends by offering to record the winner as the experiment's verdict. Say yes and the experiment closes -- the same thing the Decide button does, so the Experiments list shows it as Decided with your note and the model that won. It never records a verdict you did not ask for.
Asking for work inside an experiment
An experiment is a workspace, and the Assistant treats a request made in one as a request to work there. It knows which experiment you are in, what its question is, and how many inputs you have selected.
"Try every remove background workflow you can find" is the shape this is for. It finds the candidates, keeps the ones it can actually run, tells you the list and where each one runs, then runs all of them on the inputs you have selected -- or on everything linked into the experiment if you have selected nothing -- and ends by asking which won. "Every" means every one it can run, not a shortlist: the two-to-four limit is for a proposal it makes unprompted, not for a list you asked for.
Three things it will not do in there. It will not create a second experiment -- you are already in one. It will not run anything into the subject's own gallery. And it will not pick the winner. Ask for a different operation -- an upscale of one of these results -- and that runs in the same experiment too; you do not need a new one to change operation.
Which model it uses, and how it settles one you have not chosen
Ask for a job -- "remove the background from this", "animate shot 12" -- and the Assistant has to decide which model does it. It does not pick one out of the catalogue.
If the shot's visual style already recommends a model for that operation, that is the answer, and it says so in passing: "using RMBG, what Anime recommends for Remove Background here." It does not re-recommend, does not list alternatives, and does not ask you to confirm a decision you already made. You chose it when you set the style up, or when you decided an experiment.
If nothing answers the operation -- the shot has no style, or its style recommends nothing for it -- it proposes an experiment rather than guessing. The proposal is short: the question, the picture it would use, and two to four models it can actually run for you -- each with where it would run and, where a real run has measured one, what it costs. Models it could not run at all are left out rather than offered and then refused. Nothing runs until you say yes.
Say yes and it creates the experiment from the pictures you have selected -- they are linked into it, so the originals stay where they are -- opens it, and runs each candidate there. Every result lands in the experiment, never in your character's or shot's own gallery. Then it asks which one wins.
It will not make that call for you. Ask what it thinks and it will look at the results and tell you which it prefers -- but recording the winner is what closes the question and can make the winning model the default for a visual style, and that stays yours. See Decide for what recording it does.
It also remembers what you have already found out. A verdict you wrote on an earlier experiment -- "BiRefNet keeps the ink lines" -- is quoted back to you before anything general about a model, because what you learnt on your own artwork is better evidence than what a model is said to be good at.
Conversations
Everything you say to the Assistant lives in a conversation. The chat panel has a one-line bar across the top naming the one you are reading, and that name is also the menu: it lists your recent conversations for this project, newest first, alongside New conversation, Rename this conversation and Delete this conversation.
A new conversation names itself from your first request, so the list reads as what you were working on rather than as a row of dates. Rename it whenever you like -- a name you give it is final, and nothing overwrites it afterwards. Open the chat full screen and the same list becomes a rail down the left, where you can switch, rename and delete from one place.
Conversations are yours alone. There are no shared or team conversations.
What the Assistant remembers
The Assistant remembers a way you like to work only when you ask it to. Say it plainly: "From now on, ask me which pose before a spin", "Always use Qwen Image Edit 2511, not 2509", "Don't ask me that again". It answers with one line -- Remembered: ... -- in your own words, and from then on it follows that in every conversation and every project of yours. Nobody else's Assistant sees it.
- Only what you ask. It never keeps something because it overheard it.
- One sentence each. Short and in your words. Say "in this project" and it applies to that project only.
- Facts about your show are not memories. "Mira is half a head shorter than Venn" belongs on the character, so the Assistant tells you where it goes instead of remembering it.
- Asking again changes nothing. The same words twice are one memory.
- Forgetting. Say "forget the pose one" or "forget that". It answers Forgot: .... If your words fit more than one memory it asks which. Ask "what do you remember?" to see the list.
- A memory is advice, not a lock. If what you ask for right now says otherwise, your request wins.
- Up to twelve at a time are in front of it. If you have more, the newest twelve are.
Skills -- the specialist knowledge it reaches for
There is one Assistant, and you never have to choose who to talk to. Behind it sits a set of skills: written procedures for one kind of work each. The Assistant reads what you asked for, loads the skills that match, and follows them.
You do not name a skill and you do not turn one on. Ask "what do you see in this shot's still?" and the media-review skill is what shapes the answer; ask "which model should I use for a talking head?" and the model advice is. Several can load at once when a question spans two of them.
The skills that ship with Ordinary Animator, and what each is for:
| Skill | What it is for |
|---|---|
| Product Help | What a page is for, what goes in a field, what a good value looks like |
| Plan Guidance | What is due, what is blocking, and walking you through a criterion |
| World Building | The era, look and materials of your world, and every location and prop in it |
| Character Casting | Who your characters are -- role, want and need, arcs, relationships |
| Script Writing | Drafting outlines, scene action and dialogue, and making pasted prose parse |
| Screenplay Review | Reading a draft and saying what is wrong with it |
| Story Continuity | Checking an episode against earlier ones for contradictions and unpaid setups |
| Voice Direction | Casting a voice, fixing a bad take, pronunciation and emphasis |
| Model Advisor | Which model and build to use, whether it will run on your hardware, what it costs -- and settling an operation you have no model for with an experiment |
| Shot Design | How a scene is covered -- the shot list, framing, order, and each shot's brief |
| Shot Production | Picking what to run for a shot and filling it in, ready to go |
| Media Review | Looking at what came out, one result against its brief and a whole set for drift |
| Comfy Nodes | ComfyUI itself -- which nodes exist, how to wire them, why a workflow will not run |
| Episode Assembly | Timing, the cut, and what the episode still needs before it can be exported |
You can read this list in the app at any time: the Skills page, under Settings in the Project group of the left-hand rail, shows every skill, when the Assistant uses it, which stage of production it belongs to, and what it is allowed to do with your project. It is a reference, not a control panel -- skills are part of Ordinary Animator and there is nothing to set up, turn on or create.
A character you have written is not one of these. Skills help you make the show; the cast are in it. Talking to a character in their own voice is a different feature, on the character's page.
Watching it work
While the Assistant is working, a status line above the chat box says what it is doing right now -- reading a character, updating a prompt, looking through the workflows. On a task that takes a while it keeps updating, so you can see it is still going rather than wondering whether it has stopped.
You can keep typing while it works. The message box never locks: send one message and the cursor stays in the box, so the next thing you want to say can be written while the reply is still coming. It will not go out until the reply ends -- pressing Enter does nothing until then, and what you have written stays where it is.
If a step is taking a long time, the status line says which step and how long it has been running. That is normal for work that involves a lot of reading; if it stalls with nothing new for a long stretch, the Assistant will tell you the turn ended rather than leaving you waiting.
A long job pauses to check in. Ask for something big -- a video for every shot in a scene, a batch of pictures -- and the Assistant works through it, waiting on each render as it goes. After a good stretch of work it stops and reports: what is finished and where it is, what is still rendering on your page (that carries on by itself), and what is left. Say continue and it picks up exactly where it paused; say stop and it leaves the rest. The same pause happens, much sooner, if it finds itself repeating the same step without getting anywhere, so a stuck turn becomes a report rather than a silence. If a turn does fail outright, the apology says what had already been done before the failure, so nothing finished goes unmentioned.
Nothing changes behind your back. Every value the Assistant writes, it shows you: it takes you to the page holding that field, scrolls to it, and the field pulses blue as the new value lands. Several changes at once arrive one at a time, about a second apart, so each reads as its own beat and you can stop it part-way. This holds wherever the value lives -- a character or scene property, an operation's input on the gallery's Recipe panel, a plan item it marks done, a character it promotes to major. If the Assistant tells you it changed something, you saw it change.
Speaking instead of typing
The microphone beside the message box is voice typing. It is another way to fill the box in -- not a different Assistant, and not a conversation of its own.
- Hold the microphone button down.
- Speak. Your words appear in the message box as you say them.
- Release. The words stay in the box.
- Read what it heard, fix anything it got wrong, and press Send (or Enter) exactly as you would for a message you typed.
Because it only fills the box, voice never sends anything by surprise, and it adds to what is already there rather than replacing it -- you can type half a sentence, speak the rest, and send the lot. The reply always comes back as text; the Assistant does not speak.
Voice typing uses your browser's own speech recognition, so it needs Chrome, Edge or Safari and permission to use your microphone. In a browser without it -- Firefox -- the microphone button is greyed out and says so; everything else works as usual. The first time you hold the button your browser asks for microphone access; if you refuse, or later block it in the site's settings, holding the button says that is what happened.
If your project still lists agents
Earlier versions let you configure agents -- named specialists with their own instructions, prompts and permissions -- in the project's Agents view. Skills replaced them: they compose, so one answer can draw on two of them, and nobody has to learn a name or pick a specialist to get the right procedure.
Agents you configured before that change are still there. The project's Skills page lists them in a section of their own, below the built-in skills, and they are still yours to open, edit and delete. New ones are not created any more.
AI usage and limits
AI features (the Assistant, field help, speech generation) have an estimated cost in US dollars. Your running total for the current month appears in the left navigation, and the My Account page lists every month and, opened up, every day -- see Account.
Each account has a monthly usage limit as a safety net against accidents (a long job left running, say). The limit is yours to set: My Account shows it under Billing as the Monthly protection cap, and Change opens a box where you enter a new amount in US dollars. The limit resets at the start of each month, and so does the running total. When you reach it, AI features pause and you'll see a message saying the limit was reached and the date it resumes -- the rest of the app keeps working normally; raising the limit on My Account lets them continue.
A project you share has a second limit: the Project AI cap, which the project's owner sets on the project's Team page (it appears once the project has a member besides you). The two limits guard different money. Your limit stops the costs you pay for yourself -- asking the Assistant, writing prompts, the screenplay and plan tools -- wherever you are. The project's cap stops the project's costs, such as speech generation, for everyone on the project, the owner included; a contributor who is stopped by it is told that only the project's owner can raise it. If the owner pays a member's costs, that member's Assistant use counts against the project's cap too, and their own limit no longer applies to it. A project nobody else has joined has just one limit, yours. Voice typing itself costs nothing: the transcription happens in your browser, and what reaches the Assistant is the message you send, priced the same as one you typed.
Using the Assistant effectively
The Assistant works best when your project data is complete and well-organised. It reads character profiles, location descriptions, shot details, and screenplay text to ground its answers. Filling in Story properties on entities gives the Assistant more to work with.
For navigation: "Go to the scene where Kira and Marcus meet" works if you've written meaningful scene descriptions. Vague entity names make navigation harder.