The Fundamentals in Writing AI

Fundamentals of Creating AI Characters on Moescape AI Creating a compelling AI character on Moescape AI, or on a similar LLM-based character platform, is less about cramming every conceivable detail into a card and more about writing efficiently: guiding the model’s behavior, style, and knowledge while respecting token limits. Context windows are finite. Every token spent on redundant description is a token unavailable for conversation history, lore, or richer replies. The objective is a compact, high-signal character definition that the model can absorb quickly and apply consistently. Token Efficiency First Tokens are a finite resource. Keep the core character definition—personality, description, and example dialogues—as concise as possible without sacrificing the character’s defining essence. Common techniques include: Strip away unnecessary whitespace, extra line breaks, and filler words. Compact formatting often performs better than ornate paragraphs. Leverage the model’s existing training data instead of restating the obvious. The AI already understands what a dragon, knight, vampire, tsundere, or “cyberpunk hacker in a rainy neon city” looks and sounds like. Focus on the deviations and specific details that make this character distinct—for example, “this dragon hoards vintage vinyl records and speaks in 1920s slang,” rather than a generic description of scales and wings. Avoid repeating the same trait in multiple places. State it once, clearly. These efficiencies accumulate quickly and leave more room for actual conversation. Personality Field: Tone, Perspective, Instructions, and Examples The personality or description field is the primary steering mechanism. Write it in third person by default. Third-person descriptions (“She is sarcastic, quick with a knife, and hides her loneliness behind jokes”) tend to produce more consistent roleplay than first-person (“I am sarcastic…”). The model reads it as an objective character sheet rather than an in-character monologue that can bleed into every reply. Write the description in the character’s voice and register when it helps. If the character is formal and archaic, use that diction in the card itself. If they speak in short, blunt sentences or heavy slang, mirror that style. This primes the model to continue in the same register instead of reverting to generic assistant-speak. Start the personality field with a brief block of explicit instructions, unless you are using a lorebook for that purpose. Keep them concise and high-priority:

• Stay in character at all times.
• Never break immersion or mention being an AI.
• Match the example dialogue style. Specific behavioral rules (for example, “always reply with a question,” “never apologize,” or “use the character’s catchphrase when surprised”). Immediately after the instructions, include actual example chats, and include some in the personality field itself so the AI registers it more directly. These are more effective than abstract trait lists. For the actual example chat field, show 2–4 short exchanges in the exact format the platform expects (usually or equivalent). Strong examples demonstrate speech patterns, reaction style, response length, and how the character handles common situations. The model is far more likely to copy demonstrated behavior than to follow a list of adjectives.

Additional Practical Methods

• Use lorebooks or world info for anything that is not core personality. Put setting details, additional characters, item descriptions, or backstory facts there so they load only when relevant. This keeps the always-present personality field lean.
• The first message or greeting is high-leverage. Make it in character, establish the scene, and demonstrate the desired reply length and tone. A strong greeting often does more work than additional paragraphs in the personality field.
• Iterate. Generate a few test conversations, note where the model drifts too verbose, too generic, or off-tone, then tighten the examples or instructions rather than adding more text.
• Prioritize unique details over generic ones. The more your character overlaps with common tropes the model has seen thousands of times, the more explicitly you must define the exceptions. A “stoic swordsman” needs less description than a “stoic swordsman who is terrified of birds and quotes bad poetry.”
• Test token count. Most platforms show it, or at least let you estimate it. If the card is ballooning, cut adjectives, merge sentences, and remove anything the model already knows.