How I Use AI and Automation to Run Glitter AI

Edit: there's a new revised and expanded version of this post. You can find it here:

Read the New Version

Hey there, I'm Yuval, the CEO of Glitter AI. I don't know if you know this, but Glitter AI is a really small company. In fact, as of writing, it's just me full-time, and a couple of freelancers. As a result, I end up doing A LOT of work myself that in my previous startup, I hired people to do.

Today I'd like to take you behind-the-scenes to show you how I keep up with everything that needs to get done.

Why Automate?

The initial motivation to use AI to automate tasks came up after I launched on Product Hunt.

Glitter AI ended up winning "Product of the day." I ended up "winning" a feeling of drowning.

I had to deal with an influx of of bug reports, feature requests, and customer questions. And since I decided that this time around, I'm not going to rely on venture capital for Glitter AI, I had to get clever with limited resources.

If you’re anything like me—juggling a startup or just trying to stay on top of everything you need to do—I hope this helps you save a bit of your own sanity :)

The Initial Overwhelm After Product Hunt

Winning Product of the day was exhilirating, but overwhelming. Beyond the direct exposure on Product Hunt, Glitter AI got syndicated to a bunch of newsletters, blogs, and news outlets.

That meant that I went from practically zero traffic and zero requests on Intercom to hundreds overnight. Requests were coming in left and right: bugs, feature suggestions, questions, and a ton of general feedback. All as free-form text, with nothing categorized.

I remember sitting there, staring at the stream of tickets, thinking, "How the f*** am I going to keep up with all this?" If I didn’t figure out a way to manage this, I knew I’d either miss important details or just burn out completely.

Neither of those options seemed particularly appealing, so I got to work.

Setting Up Automations: Make.com, OpenAI, and Notion

If you haven’t used it before, Make.com is a tool that allows you to glue a bunch of APIs together to automate tasks. It’s a powerful way to set up automation pipelines that normally would have to be coded. I figured that with what large language models (LLMs) can do nowadays, I can at the very least roll something up that would summarize and categorize all the incoming requests for me.

What I had in mind was a sort of "junior-product-manager-meets-customer-support-rep" role.

Mind you, this particular automation does not help me respond to customers (that is still something I insist on doing myself), but it does help me stay on top of things.

Here’s how I set things up:

1. Intercom trigger: The majority of customer interactions at Glitter AI come through Intercom. It’s where I get bug reports, feature requests, and a lot of feedback. I needed a way to sort through all this noise and find the signal—the actionable stuff. I created an automation in Make.com that takes every incoming message from Intercom and triggers a webhook in Make.com.

2. Using OpenAI for Categorization: In Make.com, I take the conversation ID provided in the webhook, grab the full content of the conversation using the Intercom API, and then pass that on to the OpenAI API (e.g. ChatGPT) to parse and categorize the content. I set up a specific prompt that tells ChatGPT to act like an expert product manager.

   
   
   The job: summarize each conversation into clear, actionable points. Whether it’s a bug report, a feature request, or just general feedback, ChatGPT breaks it down for me.

   
   
   The prompt looks something like this:

You are an expert product manager extracting information from a conversation between admin and user.

Given the following conversation, write a summary of every bug OR feature request OR use case OR feedback provided by the user.

Some conversations may have multiple things (for example, bugs AND feature requests), but this may not always be the case.

Use your best judgment as an expert product manager to create the correct summary for this SaaS company.

Examples of use case:
"I'm using this to teach my new employees how to use our CRM"
"I want to show my VAs how to use a new system"
"I struggle with Zoom calls and this saved me the stress of jumping on one"

Examples of bugs:
"I can't log in"
"I keep getting this error"
"I get this error message when I install the desktop app"

Examples of feedback:
"I love how sleek the UX is"

Examples of feature suggestions (sometimes masked as questions):
"Does this support converting an existing video to a guide?"
"Does this integrate with Notion?"
"I wish I could remove the button from a screenshot"

3. Notion Integration: The 'Second Brain': After ChatGPT does its thing, Make.com then funnels this categorized information into Notion, where I’ve set up a database I call my “Second Brain.” Everything gets organized there—feature requests, bugs, use cases, feedback—you name it.

Notion is where I categorize each item, which is auto-linked back to the original Intercom conversation.

This is crucial because it not only keeps everything in one place, but it also helps me see how often certain features are being requested or if there’s a recurring bug that needs immediate attention. More on this in the next point.

"The beauty of this system is that it’s not just about tracking. By counting how many times a feature has been requested, I can easily prioritize what needs to be worked on next."

4. Categorizing and Counting: Once the data is in Notion, I take a quick glance to make sure everything is categorized correctly and attach the summaries to existing feature requests or bugs.
   
   The beauty of this system is that it’s not just about tracking. By counting how many times a feature has been requested, I can easily prioritize what needs to be worked on next. I use a variation of the RICE framework to decide what to prioritize.

Categorizing chaos

Before setting up this automation, I felt like I was playing whack-a-mole with customer requests. I couldn’t keep up, and that made me feel like I was letting my users down. But with this automation, I’m back in control. Now, everything is categorized, tracked, and linked back to the original conversation, and I can easily answer these questions:

1. Which features are most requested?

2. Which bugs come up the most?

3. What use cases people have for Glitter AI? (even ones I didn't think of)

4. What personas should I target in my marketing?

5. What things are unclear to users that I thought were obvious?

This not only helps me stay on top of what needs to be done but also ensures that I’m prioritizing the right things. For example, when I saw that multi-language support was a top request, I made it a priority. The same goes for the ability to export guides to different formats—another big ask from Glitter AI users.

The system also helps me quickly see if there’s a bug that’s affecting a lot of users.

Feeling like I can breathe again

More important than anything, I went from feeling anxious and overwhelmed to feeling like I was in control again. I could finally breathe again.

Setting up something like this isn't hard to do, and if you happen to be a productivity enthusiast, I can’t recommend setting up a similar system enough.

If you’re running a startup or just trying to manage a side project, I recommend you take a look at tools like Make.com, Zapier, and the likes of them. I hope it makes a world of difference to you, as it has for me.

Lastly, as a shameless plug, I will say this: not everything can be automated. Sometimes, you need to teach someone how to do something. And when that happens, please think of Glitter AI: it is a fantastic tool for creating step-by-step guides that you can share with anyone.

To record a guide, you just do your process normally, clicking and explaining what you're doing out loud.

Afterwards, you get an automatically-written shareable guide you can send to anyone, complete with screenshots and text explaining what to do.

You're welcome to give it a try at www.glitter.io :)