### Why did I build this?

1. LinkedIn and X are b2b lead goldmines. But generating content is hard. You have to figure out what to say. If you're like me, you have trouble coming up with social post ideas when staring at a blank screen. But interesting ideas arrive when I step away from my laptop and talk to people. Also, it's hard to compose the content in a way that appeases the platform algorithms.

2. I want to dogfood my product. We’re building Mixpanel for Voice AI Agents (https://voice.canonical.chat/). We map the caller journeys. We provide audio metrics like latency and interruptions. We show Voice AI devs why and when their agents are failing. I want a Voice AI agent so I can be a user of my Voice AI analytics product.

3. Voice AI will eat the world! This project is essentially a Voice UI for a social media content generation platform. We’re seeing more and more Voice AI interfaces like this (e.g., NotebookLM converts any content into a podcast), and many more use cases for Voice AI agents.

### How did I build this?

I built the Voice AI agent using Vapi (https://vapi.ai/). Vapi is amazing! It’s by far the best end-to-end Voice AI platform out there. I actually had to add latency to the Voice AI. I am a slow thinker / talker and it was unrealistic that _my_ Voice AI clone would respond at normal human intervals.

I cloned my voice using PlayHT (https://play.ht/). Yes, my voice really is that airy. I wish I could say it’s an issue with PlayHT’s voice cloning, but it’s spot on.

The hardest part has been making the social posts feel like they weren't written by an LLM. I initially was using 4o-mini, but they definitely felt written by...well...GPT. I broke the steps into smaller pieces, making one LLM call that extracts the key insights and quotes and another LLM call that generates the content. That helped, but the social posts still felt LLMy. Then I switched to using Claude. That made a big difference. I love Claude. I use Claude so much I'm not sure where I end and Claude begins.

I'd love to hear suggestions from the community on how to make the LLM-generated content feel less, as a friend of mine said, boring AF.

### Example Output

LinkedIn:

Did you know that most Voice AI agent developers are still manually listening to call recordings?

It's a time-consuming and ineffective approach to knowing what your agent is doing. It's like drinking from a fire hose.

At Canonical AI, we’re changing that by providing Voice AI developers with caller journey maps, audio metrics, and conversation analytics.

X:

Voice AI developers are currently flying blind. At Canonical AI, we're changing that by building Mixpanel for Voice AI agents.

You can see more on my recent posts on my LinkedIn (https://www.linkedin.com/in/tom-shapland-b4494212/) and Twitter (https://x.com/tom_shapland).

### Try it out!

Here's the number: +15302038359