Generating a definition, an item, an image

Three of the generator tabs make something you can review and keep: a definition, an item, and an image. This page is each one, control by control.

They live in Thalia’s dock — the ✦ Thalia button in the header, or Ctrl+I — as the Definition, Entities and Picture tabs of its strip, beside the chat. What the AI can see, and how credits work, is the AI overview. Turning a document into a board has its own walkthrough.

The shape every form shares

One form on screen at a time, filling the dock, with a pinned footer holding the Generate button. Pressing it steps the form aside and brings the chat thread forward on its own — you do not have to switch tabs. A card in the thread, in place among Thalia’s replies, narrates the run (Working…, with a progress line) and then becomes the result. The form’s tab stays in the strip: going back to it never loses what you typed, and a generation you started keeps running while you look at something else.

Along the top of each: the model picker, with its thinking toggle, and an estimate of what this run will cost (≈ N cr.). Your credit balance is not here — it is the badge in the header bar, the same one every screen shows.

Nothing reaches your project until you press Apply. Every form stages its result in that card with Apply and Discard, and a failed one offers Regenerate.

Definition

Generate a typed entity schema — fields, types, and defaults — from a description.

Control	What it is
Describe your entity definition	The prompt. Placeholder suggests Create a Character entity for a fantasy RPG with name, portrait, class, level, health points, and a biography…
@ mentions	Entities and documents. Not other definitions.
Use Tools	Off by default. Tool-calling uses fewer output tokens but more input tokens.
Quick examples	Four ready-made prompts — Character, Item, Location, Quest
Model	Lite / Std / Pro tiers, searchable, with an Extended thinking toggle

Mentioning a document is the useful part. The generator is told to read it — to take your terminology and naming conventions from the page rather than inventing its own, and to turn a list of options on it (races, classes, item types) into real enums.

Apply writes one entity definition, plus any new enums it proposed. If anything was imperfect you get Definition applied with N issue(s) rather than a silent partial write.

The result card tells you what you are about to accept: Generated: «name» and N fields · 3-column layout. If it applied traits, it says so — and warns that the traits’ own fields are not in the list above.

Entities

Generate a populated entity item from a prompt, using the definition’s fields.

Control	What it is
Entity Definition	Which shape to fill. Required.
Describe your «Definition»	The prompt, labelled with your definition’s name
@ mentions	Entities and documents
Generate portrait image	Off by default. Adds an image to the estimate.
Use project context	Off by default. Inject relevant content from your project for better results. Needs embeddings.

It reads a mentioned document for who it is about — distinguishing the main subject of a page from people merely referenced on it, and preferring a name you gave explicitly over one from the document over one invented.

You can ask for several at once. The model works out the quantity from your prompt, up to ten. The card then lets you choose which to keep: Apply (3), or Apply All (7).

Apply matches each generated field name to a real field on your definition — its own, its inherited, and its traits’ — and routes the value to the right kind of storage. A generated portrait is uploaded to your media library as «code»_portrait.png and set as the item’s picture.

Picture

Generate concept art, portraits, and scene imagery from a text prompt.

Control	What it is
Describe the image you want	The prompt. Ctrl+Enter generates.
@ mentions	None — an image prompt does not reference your project
Model	Its own searchable list of image models, defaulting to a fast one
Preset	Per model, plus Custom. Changing the model picks that model’s default preset.
Image Size	Square · Portrait · Tall · Landscape · Wide, each labelled with what it suits — or Custom, 128–4096 px, with an aspect-ratio lock
Advanced settings	Per model — style, steps, guidance, safety checker, acceleration, and so on

The card here is different: Save to Media, Download, and Discard. You can name the file before saving; it is stored as a PNG.

A picture for one character. In an entity’s picture picker, Generate opens this tab with the prompt written from that entity’s name and fields, and starts generating at once. The result’s main button then reads Set as «name»‘s picture: it saves the image to Media and sets it as the entity’s picture in one step.

Reading the result card

Every card follows the same rules, which are worth knowing once:

• Selection. Where a generation produced several things, a row of checkboxes with N of M selected, Select All and Select None.
• Warnings. A Warnings (N) strip above the body when the generation had reservations of its own.
• Apply counts. The button says Apply (3) when three of seven are selected.
• Discard clears the result. Your prompt stays, so you can adjust and go again.

What breaks

Situation	What you see
Out of credits	Insufficient AI credits. Please purchase more credits or wait for your monthly reset.
No definitions exist yet	No entity definitions available. Create one first.
Applying an item failed	Failed to apply entity items — nothing partial is written
Saving an image failed	Failed to save image, with the reason
The result card cannot render	Failed to render «result», with Clear & Try Again
A generation applied with reservations	«Thing» applied with N issue(s) — applied, and told

A character sheet, into real entities

You have a page describing a character — their stats, their kit, their history — and you want it as structured data rather than prose. There are two ways, and which you pick depends on whether the shape already exists.

If the definition exists, use the Entities tab: pick it, @-mention the document, generate. The result fills your fields, and you review it before it lands.

If it does not, you have a choice — and both happen in the same dock:

	Two steps, on the generator tabs	One turn, with Thalia
How	Definition with the document mentioned, Apply — then Entities against it, Apply	Ask Thalia for the definition and the item together
Review	Each step staged behind its own Apply	Committed as the agent finishes, then revertable
Where the fields come from	The generator reads the document directly	Thalia reads it and describes the shape

The forms do not chain: the Entities tab will not run until a definition is selected. That is the whole reason the chat agent exists as a second surface — it can carry a multi-step job, at the cost of the staging step.

Stand-ins for a playtest

Generated portraits and generated translations are how you get a project playable before the final art and the real translators arrive — a board you can read as a game, or a localized build you can test the layout of. You can replace any of it later without touching anything else: a portrait is just the item’s picture, and a translation is just that language’s text.

Translations are the AI batch translation page, a row of the Localization list that opens as a tab, listed with the rest in the AI overview; portraits are the toggle on Entities and the Picture tab above.