The Core V2 Fields

The V2 spec (still the baseline in 2026) defines these primary fields:

Field	Purpose	Notes
name	Character identifier	Used in group chats and {{char}} substitution
description	Permanent context	Always injected; keep it dense
personality	Trait summary	Often merged into description by modern cards
scenario	Situational frame	Sets the scene before the first message
first_mes	Opening message	Sent as the first assistant turn
mes_example	Example dialogue	Teaches voice and formatting via <START> blocks
creator_notes	Author metadata	Not sent to the model in most frontends

description and personality are frequently redundant. If your card defines traits twice, you are paying tokens for the same information. Consolidate.

Why mes_example Matters More Than You Think

Example dialogue is the strongest style anchor available. Models weight formatting patterns in examples heavily — line breaks, asterisk usage, and dialogue tags propagate from mes_example into live output. Three well-formed exchanges outperform a paragraph of prose describing the same voice.

V3 and the Character Book

V3 introduced character_book — an embedded lorebook. Entries trigger on keywords and inject text only when relevant.

• Keys: comma-separated triggers, case-insensitive by default
• Secondary keys: optional AND conditions
• Selective: requires both primary and secondary match
• Constant: always injected, ignoring keys
• Position: before_char or after_char controls injection order relative to the description
• Depth: how far back in context the entry is placed

This is the mechanism that makes large characters viable. A 400-token description plus a 2,000-token lorebook costs far less per turn than a 2,400-token description, because entries fire only when their keys appear.

Token Budget Reality

Every always-on field competes with chat history. A card with a bloated description pushes older messages out of the context window faster, which is the leading cause of “the bot forgot what we discussed” complaints. Treat description as a fixed tax and lorebook entries as conditional spend.

Tags, Symbols, and Formatting Conventions

Card authors use informal notation that the model interprets statistically, not literally.

• {{char}} and {{user}} — substitution macros, replaced before generation
• *asterisks* — conventionally actions or narration
• "quotes" — dialogue
• [brackets] — out-of-character or system-adjacent notes, inconsistently honored
• (parentheses) — often used for OOC remarks; also inconsistently honored
• //slashes// — italics in some Markdown renderers

None of these are enforced by SillyTavern. They are conventions the model learned from training data. A card that uses brackets for narration and asterisks for dialogue will confuse users, not the model — the model follows whatever pattern dominates mes_example.

Card Market Tags

Marketplaces layer their own taxonomy on top: female, male, nonbinary, fantasy, sci-fi, nsfw, scenario, rpg, assistant. These are discovery metadata, not model input. They never reach the prompt.

Featured Card: Archivist Lyra Vane

Lyra Vane is a useful reference card because it separates concerns cleanly:

• Description: 180 tokens covering role, tone, and constraints
• First message: opens mid-scene with a concrete task, not a greeting
• Example dialogue: four <START> blocks demonstrating dry, clipped speech
• Lorebook: 22 entries keyed to archive terminology, faction names, and location references

The card stays coherent across long sessions because the lorebook absorbs world detail that would otherwise bloat the description. Study it as a template for the description/lorebook split.

Common Field Mistakes

• Duplicating personality traits across description and personality
• Writing first_mes as a question that forces the user to invent the scene
• Using mes_example for lore instead of voice
• Setting lorebook entries to constant when a keyword would suffice
• Ignoring creator_notes, then wondering why the card ignores instructions

Conclusion

Field behavior is deterministic. Description and personality are always injected; lorebook entries are conditional; examples set voice; tags are cosmetic. Once you internalize that split, card editing becomes a budgeting exercise rather than guesswork.

MiniTavern supports the full V2/V3 field set across its iOS and Android apps, the Web Tavern, and the Chrome extension, with a Character Card Market for browsing cards that already follow these conventions. Test the Archivist Lyra Vane card there to see the description/lorebook split in practice.