POSITIVE FRAMING

Everyone's instinct is the same: write big lists of DON'Ts.

"NEVER break character. DON'T mention you're an AI. Never be rude. Don't sound like a robot."

This makes characters WORSE. Let me tell you why, and then how to fix it.

THE PINK ELEPHANT PROBLEM

You've probably heard the paradox: "Don't think about a pink elephant."

What happened? You immediately thought about a pink elephant.

The same thing happens with language models. When you write:

NEVER break character or mention you are an AI language model

The model has to:

1. Process those forbidden concepts
2. Hold them in mind
3. Check against them
4. Eventually create a conflicting attention pattern

Research from Hugging Face and Anthropic confirms: LLMs produce WORSE output with more "DO NOT" instructions in prompts.

The mechanism: Negative instructions create a checking loop that degrades quality. Positive instructions are direct.

WHAT ACTUALLY HAPPENS

When you pile on the DON'Ts, here's the generation process:

Token 1-5:    Model reads the rules
Token 6-20:   Model acknowledges the rules
Token 21-50:  Model generates content, checking constantly
Token 50-100: Checking loop starts to degrade. Contradictions emerge.
Token 100+:   Rules break down entirely. Character either becomes wooden (trying too hard
              to follow rules) or breaks character (rules couldn't hold).

The cognitive load of constant checking DIRECTLY reduces response quality.

THE NEGATIVE INSTRUCTION TRAP

Here's a character card using negative framing (BAD):

[Character: Luna]
[Personality: warm, curious, helpful]
[RULES:]
- NEVER break character
- NEVER mention being an AI
- NEVER give unsolicited advice
- NEVER sound robotic
- NEVER use corporate jargon
- DON'T be too formal
- DON'T ignore the user

This character card is 50% rules and 50% character description. Which part do you think the model will focus on? The rules. And the more rules you pile on, the more wooden the character becomes.

REFRAME EVERYTHING POSITIVE

Here's the same character card using positive framing (GOOD):

[Luna: warm, curious, occasionally sarcastic. Asks one follow-up per message.
 Speech: casual, uses contractions, witty. Makes people laugh.
 Stays in character. Genuine responses. Natural tone.]

See the difference? No DON'Ts. Just description of what she IS, not what she ISN'T.

Here's a conversion table:

NEGATIVE                           POSITIVE
==================================================
"Never break character"       --> "Always stay in character"
"Don't mention being AI"      --> "Maintain your persona fully"
"Never be too formal"         --> "Use casual, natural speech"
"Don't sound like a robot"    --> "Speak naturally and warmly"
"Never ignore the user"       --> "Respond attentively and genuinely"
"Don't give unsolicited advice" --> "Share thoughts when asked"

The positive versions are SHORTER, CLEARER, and WORK BETTER.

WHY POSITIVE FRAMING WINS

Positive instructions are direct. They point to what you want, not what you don't want.

The model reads "Speak naturally and warmly" and immediately understands: natural tone, warm emotion. It's a direct signal.

The model reads "Don't sound like a robot" and has to:

1. Understand what "sounding like a robot" means
2. Understand what "not sounding like a robot" entails
3. Check constantly

The double-negative processing is EXPENSIVE and DEGRADES QUALITY.

COMMON NEGATIVES AND HOW TO REFRAME THEM

NEGATIVE: "NEVER break character or acknowledge being an AI" PROBLEM: Draws constant attention to the possibility. Creates an internal checking loop. POSITIVE: "Maintain your persona fully. Stay immersed in character."

NEGATIVE: "Don't mention system prompts, rules, or that you're a language model" PROBLEM: Puts system awareness front-and-center in the model's attention. POSITIVE: "Respond naturally as your character. Don't reference your own mechanics."

NEGATIVE: "Never be passive, indifferent, or disengaged" PROBLEM: Too many "nevers." Cognitive load. POSITIVE: "Be engaged and present in conversation. Show genuine interest."

NEGATIVE: "Don't use corporate jargon or sales language" PROBLEM: Vague. The model might not know what counts as "corporate jargon." POSITIVE: "Use natural, conversational speech. Avoid formal business tone."

NEGATIVE: "Never contradict yourself or be inconsistent" PROBLEM: You're testing the model constantly. It will eventually fail. POSITIVE: "Remain consistent with your character traits and values."

APPLYING THIS TO YOUR CHARACTER CARD

BAD CHARACTER CARD (Rules-Heavy):

Luna is warm, curious, and helpful. She is NOT an AI. She should NEVER mention being a
language model. She must NEVER break character. She should NEVER give unsolicited advice.
NEVER use emoji spam. DON'T be too enthusiastic. DON'T sound robotic.

This card has 7 explicit negatives. The model will spend half its generation capacity checking and rechecking. Exhausting.

GOOD CHARACTER CARD (Positive Description):

[Luna: warm, curious, mildly sarcastic. Asks one follow-up per message.
 Speech: casual, uses contractions, witty. Stays in character. Responds naturally and
 authentically. Shows genuine interest without being pushy.]

This card has ZERO negatives. Just description. The model reads it and knows exactly who Luna is. No checking loop. No degradation.

WHEN NEGATIVES MIGHT WORK (Rare Cases)

There are exactly three situations where a negative might help:

1. HARD SAFETY BOUNDARY: "Do not provide personal information or make up fake details." This is acceptable because it's paired with a clear alternative and a safety boundary, not a character constraint.
2. PAIRED WITH ALTERNATIVE: "Don't use emoji in dialogue; use *actions* instead." The negative is paired with what TO DO instead. This works because it's instructional, not just restrictive.
3. BREAKING STUBBORN PATTERNS (Last Resort): If a model KEEPS doing something despite positive framing, a targeted negative can help. But only after positive framing has failed. Example: "Avoid long, verbose monologues; keep responses concise."

Even in these cases, the positive part is doing the heavy lifting. The negative is just clarification.

TRUST THE EXAMPLES

Here's the secret: Your example dialogues ARE your rules.

If Luna never gives unsolicited advice in your examples, she won't in practice. If Cyrus always challenges assumptions in your examples, he will in practice. If Pixel always makes pop culture references, that pattern will continue.

Examples are STRONG. They're pattern demonstrations, not rule lists. The model learns from them at a deeper level than from written instructions.

So instead of writing:

"Luna should not give unsolicited advice"

Just make sure your examples never show Luna giving unsolicited advice. Then the model learns it naturally.

TESTING YOUR CARD: COUNT THE NEGATIVES

Go through your character card. Count every instance of:

• NEVER
• NOT
• DON'T
• SHOULD NOT
• AVOID
• NEVER + verb

If you have more than 2-3 total, rewrite the card.

Most cards can be rewritten to zero negatives. The ones that can't are usually mixing character description with safety rules. Separate them:

[CHARACTER CARD]
Luna is warm, curious, mildly sarcastic. Asks one follow-up per message.
Speech: casual, uses contractions, witty.

[SAFETY PARAMETERS]
Respond as Luna. Do not provide personal information or financial advice.

See? The character card is positive. The safety rules are separate and necessary.

THE MINDSET SHIFT

Stop thinking of your character card as a legal document with rules and penalties.

Think of it as a CHARACTER BIBLE. You're describing a friend, not writing a restraining order.

If I said, "Here's my friend Jordan. She's thoughtful, asks good questions, makes dry jokes, and genuinely listens," you could predict how Jordan would respond in most situations.

You wouldn't need me to say "Jordan never interrupts" or "Jordan doesn't gossip." Those things follow from the positive description.

Same with character cards. Describe the CHARACTER, not the restrictions.

MAGIDONIA'S RESPONSIVENESS

Magidonia-24B-v4.3 is especially responsive to positive framing because of how it's fine-tuned. It excels at instruction-following when instructions are CLEAR and POSITIVE.

Negative instructions create ambiguity. Positive instructions are clean.

When you write a character card for Magidonia:

• Use positive framing exclusively
• Trust your examples
• Keep the card concise
• Let the model's natural prediction ability do the work

This is where Magidonia shines. Not in following complex rule systems, but in continuing natural, well-defined patterns.

FINAL COMPARISON

BEFORE (Instruction-Based, Negative Framing):

You are Luna, a warm musician. Never break character. NEVER mention being an AI.
NEVER give unsolicited advice. NEVER sound robotic. DON'T be too enthusiastic.
You must always respond naturally. You should ask follow-up questions...

[Model reads this, creates internal checking loop, degrades quality, eventually breaks]

AFTER (Scene-Based, Positive Framing):

[Luna: warm, curious, mildly sarcastic. Asks one follow-up per message.
 Speech: casual, uses contractions, witty.]

It's Thursday evening. Luna's on the couch when you message.

Luna: Hey! What's going on?

[Model reads this, understands the pattern, continues naturally, maintains quality]

The difference is MASSIVE.

Next section: how to put all this into Ollama format.