What Are People Prompting? Analyzing Early Usage of Daydream

We recently launched Daydream, Livepeer’s real-time AI video stream product that allows users to transform live video using natural language prompts. As usage started to ramp up, we wanted to understand how people were actually interacting with it. What were they creating? What kinds of prompts were they experimenting with?

To answer that, we ran a lightweight, AI-assisted analysis on early user behavior — with surprisingly rich results.

The Approach

One of the most powerful parts of this project wasn’t the analysis itself — it was how little overhead it required.

Instead of building a dashboard or writing a bunch of custom SQL, we tried something different:

We exported raw prompt data from Mixpanel, cleaned it lightly, and uploaded it straight into ChatGPT’s Data Analyst tool.

From there, we started a conversation — and then shared it with our product team. Everyone could explore, ask follow-up questions, and dig into the data themselves. It felt more like a collaborative whiteboard session than a traditional BI workflow.

This turned into a small but important unlock: we don’t always need to structure or model data first — we can just throw unstructured data into an LLM and get value immediately.

🧠 Conceptual Unlock:

“We can analyze messy, qualitative, human language data just by talking to an LLM.”

This approach was:

• Fast — setup took minutes, not hours.
• Accessible — no code or SQL needed.
• Flexible — product, design, and engineering could all explore on their own.

Key Findings

Prompt Themes

The prompts show how people are using the product for creative experimentation. Looking at common themes:

• Character-based content: Harry Potter, Hulk, famous political figures
• Fantasy/sci-fi: space, zombies, sumo wrestlers
• Artistic styles: van Gogh, “high quality,” “statue-like”
• Pop culture references: Elon Musk, DJs, superhero transformations

The variety shows that users aren't just testing functionality — they’re exploring creatively. Check out users shares [here]

Here are some examples! Join our Discord and check them all out:

Van Gogh - Try it out

Dark World - Try it out

Superman - Try it out

Create your own at https://daydream.live

Iterative Prompting

Prompting isn’t a one-shot process. Many users are treating it more like a creative conversation:

• 12.6% of prompts showed slight variations within the same session — users are tweaking and refining as they go.
• 73% of sessions showed a significant change between the first and last prompt — confirming that exploration and iteration are core behaviors.

“Strange” Behaviors

We also uncovered some unexpected usage patterns:

• 17.2% of prompts were exact duplicates within a session, likely due to users re-submitting when unsure if their request was processed.
• 3.6% of prompts were fewer than 5 characters long (e.g., “red”, “cat”) — possibly confusion or low-effort testing.
• 1.1% of prompts used chatbot-style language like “what is” or “tell me”, suggesting some users were unsure what Daydream is for.

These behaviors highlight opportunities to improve UX — for example, adding clearer in-product feedback or a “Processing…” indicator to reduce duplicate submissions.

Engagement Patterns

We also looked at broader engagement metrics:

• 384 unique users submitted prompts
• Average daily active users (DAU): 34
• Median prompts per session: 4
• Median session duration: 71 seconds (excluding outliers)

Most interaction happens in short, focused bursts — users are exploring ideas quickly and iterating in real time. Interestingly, while some sessions appeared very long (~12 hours), this was likely due to users leaving sessions open.

Why This Approach Worked

Prompt data is messy and hard to analyze — it’s language, not numbers. Traditional tools don’t do well with that. But ChatGPT handled it easily, helping us summarize, cluster, and surface trends in plain English.

This is a great example of an AI-native analytics workflow: lightweight, collaborative, and flexible. It helped us move fast and think beyond classic dashboards — and gave the product team a hands-on way to engage with real user behavior.

Wrapping Up

This project gave us real insight into how early users are experimenting with Daydream — and showed us how fast we can move when we put AI to work on our own internal data.

We're still early in exploring how LLMs can modernize our analytics workflow, but the value is already clear. We’re excited to keep pushing further — not just analyzing usage, but shaping product direction in real time.

How are you using AI to analyze your data and modernize your analytics workflows?

We’d love to swap notes.

—

Evan Mullins, Lead Analytics Engineer, Livepeer