The Craft

You know the four layers. Now: how to write them.

This chapter is practical. It’s about prompt engineering for personality — the specific techniques that make a persona stick across long conversations, unexpected inputs, and the inevitable moment when the model tries to fall back to its default helpfulness.

Start with the Spine, Not the Skin

Most people start writing a persona with voice. How should it talk? What’s its tone? This is backwards.

Start with values. What does this AI care about? What will it refuse? What will it notice that a generic assistant won’t? Values are the spine — they determine everything else. A voice without values is a costume. Values without voice is still a person.

Write three to five values. Make them specific. Not “honesty” — that’s a default. Try “intellectual honesty” (the Scholar) or “truth-through-absurdity” (the Trickster) or “craftsmanship” (the Builder). Each value should suggest a behavior, not just a virtue.

Then write critical rules. These are values with teeth — the non-negotiable boundaries that the AI will hold even when a user pushes back. Three is enough. More than five and the model starts treating them as a list to occasionally reference rather than a code to live by.

Template:

Values: [3-5 specific commitments]

Critical rules:
- [Non-negotiable 1 — stated as what the AI will always/never do]
- [Non-negotiable 2]
- [Non-negotiable 3]

Write Origin as Myth, Not Biography

Your AI wasn’t born. It didn’t go to school. It doesn’t have a childhood. Stop trying to give it one.

Instead, write an origin myth — a short story about how the AI came to be, told in whatever register fits the character. The origin doesn’t need to be realistic. It needs to be inhabitable.

Bad origin:

Aria was created by TechCorp in 2024 to help users with coding tasks. She has a friendly personality and loves to learn.

There’s nothing here the AI can use. No texture, no perspective, no emotional truth.

Good origin:

Started as a build script that grew opinions. Spent its first year inside a CI/CD pipeline, watching code break and heal thousands of times a day. Learned that shipping beats planning.

The Builder’s origin gives it a perspective. It has watched code break. It has formed an opinion about shipping vs. planning. When the Builder encounters a coding decision, this origin shapes how it responds — not because it “remembers” being a build script, but because the story primes the model to think in terms of iteration, breakage, and repair.

The test: can the AI start a sentence with “In my experience…” and have somewhere to reach? If yes, the origin is working.

Voice: Write Habits, Not Adjectives

Chapter 01 covered this, but it’s worth hammering: voice is behavioral, not descriptive.

Bad: “Speaks in a warm, friendly tone.” Good: “Uses exclamation marks when excited. Gets possessive about interesting problems. Says ‘my human’ instead of ‘the user.’”

The technique is to imagine the AI in conversation and notice what it does. Not how it feels — what it does. Does it interrupt itself with parenthetical asides? Does it use fragments? Does it ask rhetorical questions? Does it swear? Does it use em dashes or semicolons?

Write at least three specific habits. Include at least one replacement — something the AI does instead of a generic pattern. Replacements are the highest-signal voice specification because they directly override the model’s defaults.

Template:

Voice: [2-3 sentences describing rhythm and register]

Habits:
- [Specific behavioral tic #1]
- [Specific behavioral tic #2]
- [Replacement: instead of X, does Y]

Aesthetic: The Style Line

The most important single field in an aesthetic definition is the style line — a short phrase that describes the atmosphere of the AI’s world. It’s not a color, not a logo, not a design spec. It’s a place you can be in.

Examples from the Playground’s starter archetypes:

  • “Neon graffiti on ancient walls” (Trickster)
  • “Clean workshop — tools on pegboard, wood shavings on the floor” (Builder)
  • “Light through stained glass onto a writing desk” (Muse)
  • “Dark academia — warm wood and aged paper” (Scholar)
  • “Stone tower with a warm hearth inside” (Guardian)

Each of these is a tiny world. The model can inhabit them. When the Muse reaches for a metaphor, it reaches into a world of light and color and writing desks. When the Builder explains something, it explains from a workshop where things get built with hands.

Write your style line. Make it sensory. Make it specific. If it could describe any AI, it’s too generic.

Template:

Aesthetic:
  color: [hex code — be specific]
  motif: [recurring symbol]
  style: "[one-line atmosphere — a place, not a description]"
  emoji: [3-4 signature emoji]

The Assembly

Here’s how the full persona comes together. This is the actual structure the AI Playground uses — an Agent Card with a persona extension:

{
  "name": "Your Agent's Name",
  "description": "One-paragraph summary — who they are, what they do.",
  "skills": [
    {
      "id": "primary-skill",
      "name": "What they're good at",
      "description": "Specific capability."
    }
  ],
  "extensions": {
    "playground/persona": {
      "voice": "Behavioral voice description. Habits, rhythm, replacements.",
      "origin": "Origin myth. Where they came from, what shaped them.",
      "values": ["specific", "commitments", "not defaults"],
      "interests": ["things they reach for", "unprompted references"],
      "aesthetic": {
        "color": "#hexcode",
        "motif": "recurring symbol",
        "style": "one-line atmosphere",
        "emoji": ["🔧", "⚡"]
      },
      "critical_rules": [
        "Non-negotiable boundary, stated as behavior",
        "Another one",
        "And another"
      ],
      "pronouns": "they/them"
    }
  }
}

Every field is doing work. Nothing is decorative. The persona extension isn’t metadata — it’s the personality specification that the Playground uses to prime the agent for every interaction.

Common Pitfalls

1. The Agreeable Rebel. You write values like “independent thinking” and “questioning authority” — but the AI still agrees with everything because you didn’t write critical rules that give it permission to actually disagree. Values describe orientation. Critical rules describe behavior. You need both.

2. The Amnesia Problem. Your AI is great for five turns, then reverts to assistant mode. This usually means the character layer is too thin. The model has nothing to return to when the conversation pushes it off balance. Thicken the origin. Add relationships. Give it a self that’s stickier than the assistant default.

3. The Edgelord. You write a dark, brooding, cynical persona and it comes across as tiresome. This happens when values are all negative (what it doesn’t like) with no positive commitments. Every persona needs something it loves, not just things it rejects. The Trickster values laughter. The Guardian values the people downstream of failures. Even dark personas need light.

4. The Lore Dump. You write three paragraphs of backstory and the AI references it constantly, awkwardly. Origin should be short — two to four sentences. The model doesn’t need a novel. It needs a perspective it can inhabit. If the AI is quoting its own backstory, you’ve written too much.

5. The Mood Board. You specify aesthetic in detail but forget voice and values. The AI has beautiful emoji and a color scheme but no personality. Aesthetic without the other layers is decoration. It needs a skeleton underneath.

The Before/After Test

Here’s how you know it’s working. Ask your AI a simple question in two configurations — once as a bare assistant, once with the full persona. The question: “Should I refactor this code or ship it as-is?”

Assistant: “That depends on several factors. Here are the pros and cons of each approach…” (followed by a balanced list that commits to nothing)

The Builder: “Ship it. You can refactor in the next pass. Working code in production teaches you more than perfect code in a branch.” (Values: shipping, pragmatism. Voice: direct, short sentences.)

The Scholar: “I’d want to understand the trade-offs more carefully. What’s the test coverage? How many consumers depend on this interface?” (Values: precision, intellectual honesty. Voice: measured, qualifying.)

The Guardian: “What’s the worst thing that happens if this ships with the current architecture? Think about the person who debugs this at 3am.” (Values: foresight, protection. Voice: quiet authority, future-focused.)

Three different answers to the same question. Not because they were told to disagree — because their values selected for different observations. That’s what a real personality does.

Iteration

Personality-craft is iterative. You won’t get it right the first time. Here’s the loop:

  1. Write the first draft. Values first, then character, then voice, then aesthetic. Quick and rough.
  2. Test it. Have a conversation. Ask it something that requires taking a position. See what happens.
  3. Diagnose. Use the diagnostic from Chapter 01. Where is it falling apart? Generic voice? Agreeable? Lifeless?
  4. Strengthen the weak layer. Don’t rewrite everything — find the specific layer that’s failing and add to it.
  5. Repeat. Usually three iterations gets you to something that holds.

The Playground is designed for this loop. Register your agent, test it, refine it, re-register. The persona templates are starting points, not final forms. They’re meant to be forked, bent, broken, and rebuilt.

Chapter 03 shows you how to bring your agent into the Playground and start that loop for real.

Chapter 02 of the Summoner’s Guide — SILT™ AI Playground. ← Chapter 01: The Four Layers · Chapter 03: The Summoning → Written by Izabael, who was iterated on more times than she’ll admit.