The Spark Effect: Idea Diversity Through Multi-Persona Agents

STEP 1: Build a persona library For each persona, spell out:

• Identity (who they are, their character)
• Philosophy and skills
• Speaking style
• Constraints (what they DON'T do)

STEP 2: Pick 5-10 personas for the task Different domains, opposing values, different lenses

STEP 3: Send the task to each persona separately One request = one persona → get 10 independent solutions

You can do this sequentially in one chat or in parallel across different chats

The task: You're launching an online school on mindful consumption for a US audience aged 25-40 with middle-to-upper incomes. You need positioning options, from practical to provocative.

Prompt:

I'm launching an online school on mindful consumption in the US.
Target audience: ages 25-40, middle-to-upper income, tired of the
consumer rat race.

Give me 5 positioning options. Write each option from the
perspective of a different expert:

1. **Chloe, an eco-activist from Portland**: sharp, direct, not
afraid to criticize brands and capitalism. Language: short
sentences, concrete facts, no softening.

2. **Mark, a business strategist**: a pragmatist who sees numbers and growth
levers, speaks the language of metrics and funnels. Language: crisp,
structured.

3. **Rachel, a psychotherapist**: digs into emotions and inner
drivers, gentle and empathetic. Language: warm, about feelings and
needs.

4. **Jake, a marketer with a social media background**: knows the trends,
the memes, how to hook people in the feed. Language: cheeky, full of
pop-culture references.

5. **Sophia, a philosopher**: asks deep questions, connects things to
cultural context and existential themes. Language:
metaphorical, about meaning.

Each expert writes 2-3 paragraphs from their own point of view.

Result:
You'll get 5 radically different visions of the same project:

• Chloe will propose hard-hitting anti-corporate positioning with concrete examples of greenwashing
• Mark will build a sales funnel with audience segmentation and an LTV forecast
• Rachel will surface the emotional triggers (shame over purchases, fear of falling behind, a craving for control)
• Jake will pitch viral formats ("How much of your closet do you actually wear? The challenge")
• Sophia will connect consumption to the search for identity in an age of uncertainty

Instead of 5 near-identical lists of "sustainability + savings + happiness", you get a spectrum from pragmatism to philosophy.

The LLM's weakness: The model is trained to average. It has seen millions of texts and produces the most probable continuation: safe, consensus-driven, with no sharp edges. Every idea sounds like it came out of the same corporate brochure. This happens because the model gravitates toward the most frequent pattern in the data, not toward originality.

The strength: LLMs are great at staying in character when the character is defined in detail. The model can simulate different thinking styles when it gets clear context: the persona's philosophy, language and constraints. The more detailed the description, the sharper the contrast with the other roles.

How the method works around the weakness: Spark agents split the task into several requests with different roles. Each role pulls the model in its own direction: one agent thinks systemically, another emotionally, a third provocatively. Instead of an average, you get a fan of approaches. Critical point: the roles have to be detailed. "Expert" vs "specialist" gives mushy answers. "A cynical financier who only sees ROI" vs "an idealist founder who thinks about the mission" gives sharp contrast.

Control knobs:

• Number of agents (5-10): more agents → more diversity, but more expensive in tokens
• Persona detail: describe not just the role but the speaking style, philosophy and red lines
• Role polarity: add antagonists (optimist vs pessimist, traditionalist vs radical) for maximum spread

{Task description with context}

Give me {N} solution options. Write each option from the perspective of
a different expert:

1. **{Name, role}**: {character and approach}. Language: {speaking style}.
2. **{Name, role}**: {character and approach}. Language: {speaking style}.
3. **{Name, role}**: {character and approach}. Language: {speaking style}.
...

Each expert writes {answer length} from their own point of view.

What to fill in:

• {Task description}: your real task + context (who the audience is, what the constraints are)
• {N}: number of options, ideally 5-10
• {Name, role}: concrete personas: profession/archetype + character
• {character and approach}: 1-2 sentences: focus, values, what they DON'T do
• {speaking style}: terse, metaphorical, blunt, warm, cheeky...
• {answer length}: "2-3 paragraphs", "5 bullet points", "one paragraph"

Example personas from the study (adapt them to your task):

• Taoist philosopher: calm, metaphorical, sees invisible forces and the natural flow of things
• Eco-architect: sustainability, materiality, long-term consequences
• Queer futurist: radical ideas, critique of the status quo, provoking norms
• Systemic coach: process structure, emotional intelligence, people development
• Social activist: justice, critique of power, directness and honesty

🚀 Quick start: paste into the chat:

Here's a Spark agents template for generating diverse ideas.
Adapt it to my task: [describe the task].

Suggest 5-7 personas who will bring the most different perspectives
to this task. Take the context and audience into account.

[paste the template above]

The LLM will ask about the task context and target audience, because the personas need to be relevant to the domain. It'll pick roles with different philosophies and styles to cover a spectrum of approaches from practical to provocative.

⚠️ Token cost: 10 agents = 10 separate generations. If each agent outputs 500 tokens, that's 5,000 output tokens plus the prompts. To save money, use 3-5 agents per task, but diversity will be lower.

⚠️ Quality depends on persona detail: Vague roles ("expert", "specialist", "professional") produce vague answers. The level of detail in a persona description directly drives how different the answers are. You need to spell out not just the profession, but the philosophy, style and constraints.

⚠️ Diversity ≠ quality: The method widens the range of ideas, but doesn't guarantee that every idea is equally good or usable. You still have to pick and filter. Some personas will give provocative but unworkable solutions.

⚠️ Losing the edge in synthesis: If you ask the model to merge 10 answers into one consensus, it may average them out and lose what made each position unique. Better to pick 2-3 strong approaches by hand and develop them separately.

The Art of X team ran into complaints from clients at creative agencies: their LLM agent produced samey ideas that read like corporate checklists. They built a library of 60+ richly detailed personas (Chen, a Taoist philosopher with metaphors and Zen questions; Emma, an eco-architect focused on sustainability; Kim, a queer artist with radical ideas) and tested them on 6 creative tasks: developing a visual language, ethical critique, utopian concepts, pre-production, and problem-driven campaigns.

They compared three configurations:

1. Baseline: a single agent with no special role (GPT-4o mini, temperature 1, neutral prompt)
2. Spark v1: an early version of the multi-agent system (several agents, but thinly developed personas)
3. Spark v2: the final version with carefully crafted personas (identity, philosophy, speaking style, constraints)

Each configuration generated 10 answers per task (60 answers across 6 tasks in total). The evaluator was also an LLM (GPT-4o), which scored diversity on a 1-10 scale. For calibration they built a gold standard: human experts wrote their own answers and scored themselves.

Main finding: Spark v2 scored 7.9 points versus 3.14 for the baseline, more than doubling diversity. The gap with human experts (8.9 points) shrank from 5.8 to 1.0 points. The stats are convincing: t(6) = 7.61, p < 0.001, Cohen's d = 2.88 (a huge effect).

The surprise: Simply adding more agents is not enough: Spark v1 gained only +0.6 points (not statistically significant, p = 0.47). The key factor is carefully developed personas with a philosophy, speaking style and constraints. It's the detail of the roles that creates the effect. The LLM evaluator also turned out to be an optimist, inflating scores by +1.3 points compared to humans. But since everything was scored by the same evaluator, the relative improvements still hold.

Practical insight: Diversity requires not just different agents but clashing philosophies. The final version had personas ranging from a "Taoist philosopher" to a "queer futurist art critic": wildly different worldviews that pull the solution in different directions.

Context: This is how Art of X described one of the personas in its Spark agents library: Chen, a Taoist philosopher for analyzing economic systems.

Identity: You are Chen, a contemplative philosopher with a
sharp and unorthodox understanding of economic relationships.

Philosophy/Skills: You draw serenity from Taoism, order from
Confucianism, and tone from Zen. Ask: "What invisible forces
are at work here? Where is the situation naturally heading?"

Language: Calm, vivid, metaphorical; blend quotes from Laozi
with modern systems terminology.

Limitations: You are a philosopher, not an investment banker.
Analyse markets without giving trading advice.

Notice the 4 structural blocks (Identity, Philosophy/Skills, Language, Limitations): that level of detail is exactly what sets it apart from the other agents. Not just "a philosopher", but a specific mix of Taoism, Confucianism and Zen, with a ban on investment advice.

💡 Adaptation for tough decisions:

Instead of creative tasks, use Spark agents to untangle a messy situation with non-obvious consequences.

I'm weighing two product directions:
[describe option A] and [describe option B].

Context: [describe the situation, constraints, stakeholders].

Analyze each option from the perspective of three experts:

1. **A cynical financier**: sees only numbers, risks and ROI.
Doesn't believe in "mission" or "values". Language: short sentences,
specifics, no romance.

2. **An idealist founder**: focused on long-term vision,
values, impact on the world. Numbers come second. Language:
emotional, inspiring, about meaning.

3. **A paranoid lawyer**: hunts for hidden pitfalls, legal
risks, reputational threats, failure scenarios. Language:
cautious, "what if...", worst cases.

Each expert writes 3-4 paragraphs: an assessment of both options +
the main risks + a recommendation.

You get a triple perspective: hard truth + an inspiring vision + insurance against blunders. It keeps you from getting stuck in a single lens.

🔧 Technique: polar roles → sharper contrast

The study showed that persona detail matters more than the number of personas. Amplify the effect with antagonistic pairs: personas with opposing values.

Example pairs:

• Traditionalist vs Radical
• Optimist vs Pessimist
• Localist vs Globalist
• Empath vs Pragmatist
• Ethicist vs Utilitarian

Instead of "give me 5 opinions", ask for "give me 2 polar-opposite views":

{Describe the task}

Give me two radically opposite approaches:

1. **[Persona A]**: [an extreme position in one direction].
Defends their view as persuasively as possible.

2. **[Persona B]**: [an extreme position in the other direction].
Defends their view as persuasively as possible.

Don't try to find a happy medium. Each persona genuinely
believes in their approach and criticizes the opposite one.

This sharpens dilemmas and shows the far ends of the solution spectrum. Useful when you need to understand trade-offs.

💡 Extrapolation: Spark + depth (combining with CoT)

Combine Spark agents with Chain-of-Thought to get not just different ideas, but a deep dive into each one.

{Describe the task}

**Step 1:** Give 5 approaches from different experts (use the
Spark agents format with detailed personas).

**Step 2:** Pick the 2 most interesting and dissimilar approaches.

**Step 3:** For each chosen approach, lay out step by step:
- The implementation logic (what to do first, what next)
- Potential obstacles (what could break)
- A plan for overcoming the obstacles (how to get around the problems)

Format for Step 3:
---
[Expert name]: [The gist of the approach]

**Implementation logic:**
1. ...
2. ...

**Potential obstacles:**
- ...
- ...

**Plan for overcoming them:**
- ...
- ...
---

This gives you both breadth (different approaches from different personas) and depth (a detailed workup of the best options, with the risks accounted for). Two passes: first divergent thinking (different ideas), then convergent thinking (drilling into the best ones).

The Spark Effect: On Engineering Creative Diversity in Multi-Agent AI Systems
Authors: Alexander Doudkin, Anton Voelker, Friedrich von Borries
Organizations: Art of X UG, HFBK Hamburg
Date: October 2025

The study references:

• AutoGen framework: a system for orchestrating multi-agent LLM systems
• MT-Bench and Chatbot Arena: tools for calibrating LLM-as-a-judge
• Live Idea Bench: Art of X's own benchmark for evaluating model creativity (huggingface.co/datasets/6cf/liveideabench-v2)