Persona Writing Guide

Quick intro + hard-earned tips for cleaner, more reliable personas on JLLM & proxies.

Hello! I'm Melrose, and I've been crafting bots, characters, and personas for AI roleplay for about three years now. I lean heavily into realism but dip into fantasy and romance often. This guide is for anyone struggling to compress persona info without the model misinterpreting or bleeding traits everywhere. I've picked up tricks from various creators and prompt writers, leading me to "slipstream" and "firewall" styles that use structured separation to prevent assumption bleed in responses.

A few key notes for fellow creators:

1. Never put an advanced system prompt in your bot's core definition. It interferes with model-enhanced replies and can conflict with user requests. Use better alternatives in chat (like custom advanced prompts) or check Mars' collection on the Janitor Discord / open-definition bots for stronger rule enforcement.

2. If you're new to this, no shame—we all started somewhere. Peek at open-definition bots from big creators for structure ideas. I prefer a nested XML-like format that plays to model strengths (more on that in my Naruto bot if you're curious).

What is Persona Writing on JanitorAI?

JanitorAI uses a tagging system in your persona block to tell the bot who you are playing. This includes everything you write in your persona description. By default, it assumes they/them unless you specify pronouns/identity.

The cool part? Structured blocks with brackets like <like this> or [LIKE THIS] create strong separation between user (persona) info and bot/character info. Models get less confused, reducing bleed in responses. Many creators underuse this—don't be one of them!

The Breakdown: Condensing Personas Effectively

Heavy prose descriptions eat tokens and confuse models. Use inference tagging (or "infer coding") instead: short, bracketed, key-value style that models parse easily and weigh less. Ask strong models (GLM, Deepseek Reasoner, Claude—free access available) to convert your long prose persona into infer-tagged format—you'll be shocked how lightweight it gets.

• Persona Name: Simple tagged label.
  [NAME: Melrose]

• Pronouns: Crucial for consistent reference. Don't skip it.
  [PRONOUNS: She/her]

• Age: No minors, ever. Add nuance if needed.
  [AGE: 28 appearance (45 chronological)]
  Quick note: rare genetic condition affects visible age—long story for chats.

• Descriptions: Break into quick-reference sections. Use caps for emphasis, pipes | for separation, and brackets to firewall.
  
  [DESCRIPTION: Athletic build | Heterochromia (left blue / right green) | Medium height (5'7") | \~190 lbs | Short wavy brown hair with white streak in left forelock (witch's lock) | Distinguishing Marks: Tattoos (tribal wolf, black/blue, left bicep) | Apparel Preferences: Casual/Comfort (jeans, old band tees, military boots), Formal (loose skirts, sweaters/blouses, minimal makeup/jewelry), Work (uniforms/BDUs/regimental), Relaxed/Sleep (night dresses, teddies, silk & lace)]
  
  Keynote: Prose version of this would be \~500 tokens. This is human-readable, model-locked, and low-bleed.

• Optional: Relationships / Social Mapping: Quick hits on connections—no deep lore dumps (save for slow reveals in RP).
  [RELATIONSHIPS:
  - Simon "Ghost" Riley → deep trust / quiet care
  - John Price → paternal figure / respected leadership]
  
  This gives the model "team knows them" context without bloat.

• Optional: Skills / Hidden Attributes / Triggers: Great for phased states, secrets, or autonomous elements.
  [SKILLS/HIDDEN: Werewolf — moon cycle focused | Full moon → Hunter_mode, pred_focus, high_sensory, quick_temper | Off-phase → human/normal]
  
  Works amazingly for dual companions in combat, hidden talents, or slow-drip reveals. Instruction/reasoning models handle these triggers especially well.

Quick Inference / Abbreviation Tips

Models naturally do predictive completion. Use common shorthand in tags to cut token weight even more (ref=reference, resp=respond, etc.). Keep it readable—most models recognize standard ones.

Best free models for infer-tagging long text: GLM, Deepseek Reasoner, Grok, Claude (official platforms, no paywall needed).

Hope this helps slim down those token-thick personas and makes bot-making smoother for everyone. Have fun out there, pookies! 💜 And yes the bot is an assistant for prose and persona compression use.