AI character card design: drafting with LLMs, then making it safe to ship
People who search ai card design or AI character card design often want a workflow: use an LLM to speed up writing personality, greetings, and lorebook entries, then tighten the result so Silly Tavern and downstream models behave. This note is that workflow—not a prompt pack, and not official Silly Tavern guidance.
What the LLM is good at
- Outlining sections: bullet lists of traits, sample beats for
mes_example, alternate openers. - Reformatting text you already wrote into clearer paragraphs.
- Spotting contradictions between
scenarioandcharacter_bookentries—if you paste both.
What the LLM is bad at (watch these)
- Exact JSON keys and extension fields: models hallucinate key names or nest objects wrong. Always compare against the field series or your exporter’s output.
- Your ST version’s behavior: “works in chat” and “valid file” are different bars—validate in a real editor / import cycle.
- Privacy: pasting a full card into a cloud chat uploads your scenario and lore—use policies you are comfortable with, or work locally.
A practical AI card design loop
- Draft voice in plain language; ask the model for short
personality-style bullets, then you edit for tone. - Draft openings as separate chunks; pick one
first_mes, park others inalternate_greetingsif appropriate. - Lorebook last: add entries only for facts that actually need retrieval triggers—character_book explains structure.
- Ship check: open the card in Silly Tavern or a trusted editor, run a few real chats, revise.
Silly Tavern extensions that call an LLM inside ST are covered in (5/5) Extensions & art—different from “chat with ChatGPT in a browser,” but the same review rule applies.
Design and tools
- Character card design (information architecture): Character card design.
- macOS editing (files on disk): macOS Silly Tavern card edit workflow.
See also
Browse, preview, and edit embedded JSON in PNGs locally on Mac with Sillycard.