What Makes A Good Persona?

🎭What Makes a Good Persona?

Think of your persona like packing for a trip - you want to bring just enough to be prepared, but not so much that your suitcase bursts open. The sweet spot? Around 500 tokens or less. Any more than that and you're just weighing yourself down with unnecessary baggage that the AI will either ignore or misuse.

A persona crammed with every possible detail doesn't make the AI understand you better - it just gives it more opportunities to make wrong assumptions. When you include personality traits or backstory in your persona, you're essentially handing the AI a script of how it should roleplay you. Before you know it, you'll find the bot narrating your actions ("{{user}} rolls their eyes dramatically") or making decisions for your character.

Why XML Tags Help

XML tags act like clear signposts for the AI. When you list traits without structure, the AI may absorb them as general knowledge rather than your specific traits. XML creates a container that says: "This is descriptive information about {{user}}". Cleanly tagged personas are also easier for the AI to parse, reducing "token bleed" (where the model wastes processing power on poorly organized info).

What Actually Matters

Focus on the concrete, unchanging facts that help the AI visualize you without boxing you in:

<user_persona>Alex:
- average height, lanky build; young-adult
- male
- sun-bleached hair
- blue eyes
- smells like cookies</user_persona>

This gives the model enough to work with, while still leaving room for your character to develop naturally. Notice how we:

• Use comparisons like “average height” instead of exact numbers
• Describe how something looks, rather than telling the AI what to do with it
• Avoid controlling behaviour and stick to appearance and vibes

If you notice the bot saying something odd about your persona, like treating “sun-bleached hair” as if your head went through a tanning bed, it’s a clue your description needs refining. Ambiguity invites hallucination.

For example:

• If the AI thinks "sun-bleached" means blonde and surfer-y, but your character’s not that, clarify it. ("lightened from black to copper-blonde by sun exposure")
• If it keeps describing you as lanky when that’s not the vibe, say "toned but not overly muscular" or "solid build, not lanky."

This becomes especially important with more nuanced or complex personas—like those with traits that aren’t stereotypical.

Case Study: an actual persona

Let’s break down a real example:

<user_persona>Faust:
- 6'1" tall(long|lanky frame); young-adult
- male(they/them); male anatomy(penis|flat chest)
- long hair(base=black|streaks=orange,white)
- green eyes
- smells like vanilla(not-parfume)</user_persona>

Faust is nonbinary, but presents in a very male-oriented way. And here’s the thing: JLLM (and many LLMs) often struggle with the difference between gender identity, presentation, and anatomy. Without specific clarifications, the AI will make incorrect assumptions, especially in NSFW or emotionally nuanced scenes.

To fix that, I:

• Clarified gender and pronouns: "male(they/them)" tells the AI they present male but use nonbinary pronouns. Reinforced by writing in third person.
• Defined anatomy explicitly: so there's no confusion about what the bot should expect physically.

Next Up: Chat Memory & Context Management

Updated on: 31/07/2025