The AI Creator
A conversational, self-healing card builder: you talk to a "character architect" model, and a live Sheet fills in as you go — no "Generate now" button to press.
What it is
The Creator turns a description into a finished chara_card_v2 card by conversation. You tell an architect model what you want, in any phrasing or language, and it builds the card with you. As it works, a live Sheet — the card's data block — fills in section by section so you can watch the card take shape.
Each in-progress card is its own session. You can keep several going at once, switch between them, and come back later; nothing is generated until you describe what you want, and nothing is saved to your library until you choose to save it.
The freeform flow
There is one way to build a card, and it's freeform. (An older guided flow that emitted one section and paused for confirmation has been removed.)
- Describe your intent — a vibe, a premise, a name, a few traits — in whatever words you like.
- The architect emits the card and starts populating the Sheet.
- A convergent loop then runs until the card is complete, filling in and tightening every section.
You describe the character once and get a finished, full-depth card without babysitting each field.
Completeness-driven, not turn-counted
The loop doesn't stop after a fixed number of turns — it stops when the card is genuinely complete. A pure-function engine checks every required field for three things: that it's present, that it clears a per-mode length floor, and (for prose fields) that it terminated cleanly rather than being cut off mid-sentence. The loop keeps going until those checks pass, self-correcting against truncation and skipped fields along the way.
That convergence is bounded so it can never run away:
| Guardrail | What it does |
|---|---|
| Stall cap | After a set number of no-progress turns in a row, the loop stops instead of grinding. |
| Iteration ceiling | A hard maximum on total passes, so a session always terminates. |
| Persisted "exhausted" flag | If a session gives up (e.g. a refusing provider), reopening it won't silently burn a fresh paid-retry budget on the same wall. |
Dedicated review pass
After the card is assembled, a final QA reviewer audits it. It either confirms the card is clean or emits targeted corrections that re-do only the broken fields — then re-reviews. It's a built-in editor that catches the model's own mistakes before you save, running in bounded rounds.
The three modes
The Creator can build different kinds of card.
Character architect
Builds a full character card: appearance, psychology, scenario, opening message, example dialogue, tags, and the rest of the chara_card_v2 fields. This is the default mode for making one character.
Scenario architect
Builds an omniscient narrator / Gamemaster card that voices NPCs inside a setting, rather than a single character. Its Description is assembled from structured XML sections and merged section by section:
<Narrator>,<Reading the Persona>,<Scene Setup>,<Tone><World>— the setting itself<NPCs>— a real cast, each built by a per-NPC mini-architect
Because the world and the cast are required for the card to count as complete, scenario mode produces a playable setting with characters in it, not just one figure.
Personas
The Creator also helps you build a persona — a profile for who you are in a chat. (See Characters & personas for how personas are used at chat time.)
Vision and image references
Attach a reference image and a vision model writes a clinical appearance profile that feeds the card. This runs at a low temperature on purpose, so the description stays faithful to the image instead of getting creative. Large images are auto-downscaled before upload, to save time and per-tile billing.
Attach cards and documents
You can seed or inform a build with existing material:
- An existing card — drop in a
chara_cardand its full JSON is fed to the architect as context, so you can remix or riff on it. - A document — attach a PDF or text file and its text is extracted (PDF via Syncfusion) so you can base a character on source material.
Edit-with-AI mode
Open an existing character straight into a Creator session and make freeform edits by conversation. Saving updates that character in place; a save-as-copy option lets you fork it into a new card instead of overwriting the original. It's the same architect, pointed at a card you already have.
Forkable architect prompts (Creator Presets)
The architect's behavior is itself a prompt you can fork. The base prompts (character, scenario, edit) ship as a locked "Pyre Default" preset — copy it to get an editable clone you can tune.
On top of that, a free-form "Your additions" addendum is appended to every architect prompt — a place for standing instructions like "always reply in Portuguese" or "always include at least three NPCs" without touching the underlying machinery.
Per-creator-call sampling
The Creator's stages have different jobs, so each gets its own sampling profile. There are separate temperature and token settings for the design chat, the vision analysis, and the canvas updater that writes the Sheet — so the chat can be creative, the vision pass faithful, and the sheet writer near-deterministic.
Sessions
Sessions are managed in a drawer:
- Auto-named from the card's name as it fills in.
- Pin, rename, or delete any session.
- Retry a turn to restore the pre-turn Sheet snapshot and re-run.
- Empty-session auto-cleanup keeps the drawer tidy.
- A "saved" badge marks sessions you've exported to your library.
A recommended-models banner points you at models known to do this well, so you don't burn hours on a bad fit before realizing why.
The model matters more than the prompt
Card generation is a multi-pass loop, so it takes time — that's the convergence and review passes doing their job. And for the Creator specifically, the model matters more than the prompt: an uncensored or DeepSeek-family model tends to give the best results. Recommendations live in-app under More → Character Creator.
See also
- Characters & personas — the
chara_card_v2data model, and how personas work at chat time. - Chat presets & sampling — forkable prompt presets and the SillyTavern-style sampling overrides.
- Providers & API connections — pick the model the Creator runs on, and configure your keys.