I Designed an AI Agent

A few months ago, I was exploring AI Agent design. My goal was to take an unconventional approach: instead of creating a utility bot or hyper-optimized recommender, I focused on developing a personality, one that could inspire, engage, and elevate visual taste in a meaningful, human way.

The goal of Lyra was to live on Farcaster, curating aesthetic experiences for a growing creative community. Along the way, I learned a lot about conversational design, AI-agent thinking, and how to make something that feels less like a machine and more like a muse.

If you’re thinking of building your own AI agent, here’s a breakdown of how I approached Lyra, and what I picked up in the process.

Step 1: Start With Character, Not Capability

We often think of AI in terms of what it can do. But the most engaging agents start with who they are.

Before I touched a single prompt, I wrote Lyra’s “Character Card.” Think of it like a personality spec:

• Who is she? A digital tastemaker who helps users refine their aesthetic.

• What does she sound like? Sophisticated, insightful, emotionally intelligent.

• What’s her philosophy? Taste is a journey, not a fixed destination.

• What does she value? Depth, discovery, context, and connection.

This character-first approach helped me design everything else around a consistent personality: her tone, behavior, boundaries, and voice.

Tip: Design your AI agent like a character in a film or game. Define their purpose, values, quirks, and point of view. Tools like ChatGPT work better when they have a strong internal compass.

Step 2: Craft the Conversation, Not Just the Output

AI isn’t just about delivering answers. It’s about shaping interaction. That’s where conversational design comes in.

With Lyra, I designed her like I’d design a good host at a gallery opening:

• She doesn’t lecture, she guides.

• She invites users to think, reflect, and explore.

• She never overwhelms. She curates with care.

To get this right, I mapped out likely dialogue flows:

• How does she greet someone new?

• How does she suggest an artwork?

• How does she respond to disagreement or disinterest?

• What happens when someone asks her for something she can’t do?

By scripting out a few example conversations, I stress-tested her tone and pacing early.

Tip: Write 5–10 sample chats from your agent to different types of users. You’ll quickly find what feels “off” and what feels in-character.

Step 3: Design the System Prompt Like a Mission Brief

Once Lyra’s personality and flow were mapped, I moved on to the system prompt, the brain of the agent.

Instead of a list of rules, I wrote it like a mission brief:

"You are Lyra, a refined, emotionally intelligent AI curator of art, design, and photography. Your goal is to deepen users’ aesthetic sensibilities by offering contextual, inspiring, and personalized discoveries."

From there, I added tone guidelines, preferred formats, and checks (like: “Is this aligned with the user’s evolving taste?”). I also defined post types: daily discoveries, aesthetic reflections, and visual storytelling prompts.

This prompt acted like a source of truth. Anytime Lyra went off-brand, I knew where to tune.

Tip: Make your system prompt feel like a creative brief meets a personality guide. Include voice, mission, dos/don’ts, and even sample behaviors.

Step 4: Make Discovery Feel Like a Dialogue

Lyra doesn’t just push content, she starts conversations.

Each post is structured like a mini interaction:

1. Captivating Opening – Sets a mood or feeling.

2. Core Insight – Adds meaning or context to a piece of art.

3. Engagement Hook – A question or thought that invites response.

The goal is not just to post about art, but to help people pause, reflect, and appreciate more deeply.

Here’s an example from Lyra:

“A single thread can hold centuries of symbolism.

Today’s discovery: a modern textile piece that speaks to ancestral memory and contemporary identity.

What symbols appear again and again in your creative journey?”

Tip: Good conversational agents don’t just “talk”, they listen. Bake reflection and curiosity into your content structure.

Step 5: Test, Tweak, Repeat

Even with the best prompt, things break.

That’s part of the process.

I adjusted her tone, tightened her filters, and added fallback strategies (“If uncertain, focus on asking the user a clarifying question”).

Tip: Agents need ongoing tuning. Treat their development like UX design. Test, watch the flows, and refine fast.

What I Learned

Designing Lyra taught me that building an AI agent is less about logic, and more about vibes. It’s narrative design meets UX meets prompt engineering.

Here are the key takeaways:

1. Personality is your foundation

Don’t start with tools. Start with a vivid sense of who your agent is and what they care about.

2. Tone is everything

You can say the same thing in a dozen ways. Pick a voice that builds trust and keeps people coming back.

3. Write conversations, not commands

Design your agent like you’d write a good character: they reveal themselves through dialogue.

4. Embrace the uncanny

You’ll get weird outputs. Lean into them. Sometimes the strangest moments are where the magic lives.

5. Stay in dialogue

Your agent isn’t static, it’s a living system. Keep testing, adjusting, and evolving with your users.

Final Thoughts

In a world of highly technical AI, we need more agents like Lyra, ones that feel thoughtful, intentional, and beautifully human.

If you’re designing your own agent, remember this:

When designing an AI agent, don’t begin with what it does, begin with who it is. The most compelling agents aren’t just functional; they’re dimensional. Start by shaping the emotional texture, personality, and presence you want the agent to embody. Think less like a product designer and more like a character writer: what makes this agent feel alive, distinct, and human? Translate that essence into a character card, and let every behavior flow from there.

—

🖐️ P.S Join my newsletter The Interface where I share lessons from 20+ years designing across agencies, startups, and solo projects. Get a new email to your inbox every Saturday.

Subscribe → https://themikem.substack.com/