I published my first post on Substack on December 26, 2022. I posted for 20 weeks straight before stopping completely when my post about being laid off from Meta had 10x impressions than my average post. I didn’t publish again for another year and a half.

I came back in 2025, nearly three years to the day that I had first started on Substack. I was working on launching my coaching business and wanted a way to connect with my audience. I also wanted to be better at writing. I posted nearly once a week to once a month for 10 months.

In November 2025, I started to be more active and make connections with the Substack community. My feed filled with thought-provoking, poignant, funny, creative posts, many of them related to AI and many of them written by women.

Now, almost six months to the day I changed how I showed up on Substack, I’m just shy of 500 subscribers on my main publication, Step Up Step Together.

When I reach a milestone, whether it’s on Substack or earning my PhD, my mind immediately goes to the many people who were part of that journey. Maybe they gave direct support or advice. Sometimes it was an engaging conversation that taught me a new way to think about the human experience.

So, for this 500 subscriber milestone, I built an app to celebrate all the people who have connected with me on Substack.

Check out First Replies here!

Here’s how I built First Replies.

Step 1: Use what you have

When I started this project, all I knew was — I wanted to highlight my connections on Substack, specifically anyone who has commented on my posts or replied to my comments to them.

Substack doesn’t easily provide this data, but luckily, about a month and a half ago, I built an app called Substack Replies, which surfaces notes or comments that I haven’t responded to.

The app makes API calls to Substack to retrieve my Substack conversation data and then saves it in a local database on my computer.

So I went to Claude Code, and with no preamble, said:

“in src/substack_replies [the project directory], there’s a db [database], I want to know who my top commenters are, like top 50”

It beep boop beeped, ran SQL queries, and returned them.

Then the fun magic happened.

As one idea came to life, another idea popped up.

What if I showcased my commenters on the subscriber growth graph?

Step 2: Build the visualization you want from the best data you have

I took a screenshot of the subscriber graph and asked Claude Code to recreate it. I expected it to estimate the values based on the graph gridlines (because it had done this for me before. …I just realized, it wasn’t Claude Code that did it. It was Claude.)

“are you able to recreate this graph [Image #1]”

It beep booped again, and basically said no. It wanted raw data.

I thought about pushing on it, but I had also just exported my official Substack data. So I gave that data to Claude Code, and it recreated the graph beautifully.

Then, I turned up the heat.

“looks pretty good!

Can we turn this into an interactive webpage where there are dots for the top 50 commenters by the date of their first comment, and when you hover over it, it shows their name?”

And Claude Code did it. (Good job Claude Code.)

Step 3: Appreciate the unexpected moments of AI

Then I asked Claude Code to do this:

“when i hover over a pin, really pop it out, like move out with big space, right now it’s just getting bigger in place.

also make the pop out happen a little faster”

Claude Code complied. When I saw what it created, I couldn’t stop laughing. I had to share it with Substack:

I really wanted to keep the ridiculous bouncing, but it was too hard to click on the bubbles and see the commenters, so I removed it. Sniff, sniff.

Step 4: A compilation of requests

After that, I had a normal back and forth conversation with Claude Code. To give you an idea of how I work with my AI, here is a selection of the requests I made, shown exactly how I asked them.

On bubble actions

On stem appearance

On taking up space

Step 5: Check your approach

To build the app, Claude Code used HTML, which is one of the simpler coding approaches. We were having some issues with animations and the images, so I asked if we should use a different tech stack.

“Also shoudl this be written in something else like node.js? or whatever else you normally use?”

It told me no:

“Plain HTML + D3 is actually the right tool here. Node.js would just add a server you don’t need. If you wanted live data (auto-refreshing from the DB), a small Python script that regenerates the JSON would be cleaner than a server.”

And that was a good enough answer to me.

The other advantage of keeping the app in HTML is that it can be deployed through GitHub Pages for free. It also means that when you make a code change and merge it into main (the live version), the Page will automatically update to reflect those changes.

Step 6: Track your changes — Here comes GitHub! (+ privacy considerations)

After I had a good working version, I wanted to try a drastic change. Instead of bubbles linking to Substack profiles, I wanted them to link to the first comment a commenter had said. But if I decided I didn’t like the change, I wanted to be able to easily revert to the original version.

So I asked Claude Code to create a GitHub repo for me.

I asked Claude Code:

“can you create a gitrepo for this, just local for now so we can get this under version control. Put it in src and name it alyssa-subscribers, include a .gitignore”

I made the repo local so that I could vet it in a safe environment. Once I update my changes to the remote repo, if the repo is public, then anyone could see my changes.

Step 7: Appreciate what you’ve made

When I tested the change of including the first comments people made in the bubbles, that’s when the impact of this project hit me. It took my breath away.

Highlighting over every bubble filled me with appreciation for these incredible people on Substack and the connections that we’ve made.

For some, I’ve had over 100 back and forth comments with them! For others, maybe only 2 or 3. But even the ones where there were fewer interactions, I remember them, and they felt meaningful.

There are over 300 people represented in this dashboard. 300 people where either you or me decided — we want to share what we’re thinking with each other. And every single one of these people have shown me how digital spaces can be welcoming spaces, not lonely, isolating ones.

I shared that sentiment with Claude Code, and it noted as well:

Step Epilogue: Know when to stop (that time is not now)

I couldn’t get the bouncing bubbles out of my brain. And… my little orange was missing. I couldn’t ship this app without mitigating both travesties.

I asked Claude Code:

“Also do you remember that mode where the bubbles popped up and down really fast? Are youa ble to make that a toggle mode too, like ridiculous mode?

And it did. I was so happy with it, I made the ridiculous mode the default mode.

And then I gave it photos of my little oranges and asked it to add them to the graph, including a way for users to interact with them. When I tested it out, I shrieked with delight, then couldn’t stop laughing.

Bounce, oranges, bounce!

To check out First Replies, bouncing bubbles, and multiplying bouncing oranges here:

Check out First Replies here!

And if you’re interested in checking out the GitHub repo, here it is:

Check out the First Replies GitHub repo!

Thank you for being here

Thank you each and every one of you nearly 500 subscribers (and followers and watchers!) for being here.

For me, number of subscribers is not a number to optimize or grow to no end.

When I see my subscriber count, I think —

One thing that surprised me as I grew was how much each new subscriber was often someone I knew or had interacted with. It came from making a connection with another person. Even if the interaction was brief, it’s meaningful that two people in a world full of 8 billion connected over an idea, a joke, or a thought-provoking comment.

I am inspired by you, Substack, and I’m grateful for the time we spend together. I have loved finding joy in the little things with you and how we are working to make our way through the messiness of being human in this new technology world together.

We can use AI to celebrating being human.
Join me so we can find creative ways to do it together.

Subscribe now

Share

Want more? Here are other posts on how you can build with AI without losing yourself:

The Off Hours 🌙

How We Turned Our Substack Brands into Songs using Suno AI

TL;DR We used AI to turn our brands into actual songs. In this post I walk you through exactly how I did it — and how you can build your own brand soundtrack with Suno.ai, even if you've never touched music software. I built the songwriting process. Pinkie made the album cover. Both prompts are yours to use…

Read more

10 days ago · 7 likes · 3 comments · Dallas Payne

That was my mom a few months ago. We were driving, and she was sandwiched between my two kids in the back of the seat. And we were talking about AI.

Fast forward, and we just finished recording a video of us using AI together. Not just using AI, but building with it. My mom and I were giving commands to Claude Code, and answering its questions.

And it wasn’t with just my mom.

We brought my almost 8 year old daughter into the mix. In fact, once I asked my daughter if she wanted to be part of my community AI project, that’s all she wanted to do that morning.

AI means so many different things to different people. Even within the same person, from one moment to the next, AI can be exciting or scary. There are many consequential issues that need to be addressed related to AI whether it be its biased training and outputs to its environmental impact.

But today, AI was a connective tool for me and my family. It was fun.

In a million years, I never would have dreamt I would be staring at a command line with my mother AND my daughter in the same day at the same time.

But we did it. And the feeling we walked away with at the end of it was that this is awesome. This technology is awesome.

To some, they may think that all we did was put a simple picture on a webpage.

To us, this is just the first step in being able to create a vision, a product that connects to other people and that we never would have been able to build before without AI.

In this post, I’ll share

• The nuts and bolts of how my daughter and I built her room. Think of this as the written down version of the tutorial.

• How we leveled up the room with my mother, focusing more on her ah-ha moments than the technical details

• What this means for you and how working on AI can be fun instead of urgent.

Here’s what we did.

Part 1: Why not make a video tutorial together?

I was working on a documentation tutorial for Open Room, a community AI project I created to help people learn how to build with AI together and create something fun and connective in the process.

I was planning to create a video tutorial to for with it, my first one ever on Substack.

My mom also happened to be visiting. She lives two hours away, so she’s here every few weeks.

When she’s here, we stay up late with my husband talking about everything and anything. Recently, AI has been a common topic. And not in a heated, contentious, “we’re debating” type way. But in a deep, thoughtful, “we’re all trying to figure this out together” kind of way.

As we talked, an idea bubbled up inside me. What if, instead of me running through the tutorial by myself, I did it with my mother… and brought my daughter along, too?

Working on AI with my daughter is actually not a far stretch. We’ve used AI together several times, whether it was to create claymation versions of our selves, bedtime stories, or vibe code mystical creature versions of the game 2048.

When it came to building with AI with my mother? That was new territory.

But I had intentionally built Open Room to be accessible to anyone, usable by anyone. The whole idea was to create something that was delightful for you — a room that represented something about who you are and that you could use to connect with others — but also to serve as a pedagogical tool.

It answers the question — How can you teach someone how to build with AI in a way that’s easy, approachable, and fun and still learn technical best practices?

I had a perfect opportunity to find out.

Part 2: Building a room with my daughter

I started first with my daughter. I spent the morning going through the workflow with her so we would know what to anticipate when we were in front of the camera.

Here I’ll share with the steps we went through, in detail, to build her room. Feel free to follow along to build your own room!

Step 1: Reserve a room

The first step we navigated to the site: https://open-room-open-source.vercel.app/.

Once there, Kaylee reserved a room by clicking “+ Add room” and submitting a GitHub username (one of my alternates) and an email address.

After reserving the room, a confirmation page pops up.

Every room is assigned a registry id — a two word compound — and that id is used to track information about the room across the site and as changes are made to the site. This is different from the room’s name, which users can control and change.

Then, we started to talk to Claude Code.

Step 2: Get set up with Claude Code

After reserving the room, we opened up Claude Code. I told Kaylee to give it this prompt — basically she provided a link to where the code for the site is stored and a request:

“https://github.com/alyssafuward/open-room-open-source I need help setting up my room”

The first thing Claude Code will do is navigate to the website and read it.

When I set this project up, I knew Claude Code would know what to do with this link. It would know how to read a GitHub repository (see Terminology Check below). And within the repository, I had built in instructions for Claude Code to follow so it could guide the builder.

Terminology check: GitHub is a cloud platform that allows users to store code, track changes, and collaborate together on software projects. The ability to track code changes is crucial when building technical products, and that still applies with AI.

A GitHub repository is a storage space for all the files and tracked changes associated with a given project. It can also include the issues made toward that project. It can also be called a repo.

“You mean like a house?” Kaylee asked me. Yes, actually, Kaylee, it’s very much like a house!

After reading the setup instructions, Claude Code then knew to ask Kaylee:

“What’s your GitHub username? I’ll look up your reserved room.”

And she told it: “afwcoaching-beep.”

What Claude Code then finds information about Kaylee’s room based on the GitHub user name.

Here’s how it does it:

When you make a reservation on the site, the app actually makes what’s called a GitHub issue in GitHub. It’s also called a ticket. It’s basically just a way to say — there’s a change I want to make to the site, and here is information about it.

For Open Room, this reservation ticket includes metadata information

And this is where the power of AI comes in to connect the dots.

When you tell Claude Code “this is my username”, I programmed it to look at the GitHub issues and find the one labeled with your user name. That ticket that contains information that becomes context Claude Code can use to help you build your room.

And indeed, Claude Code found the GitHub issue associated with Kaylee’s GitHub username and confirmed with her that it was correct.

“Your room ID is quiet-glade (issue #111). Does that sound right to you?”

After saying “yes”, Claude Code then checks to see if the base repository had been forked yet — where it’s been copied.

Terminology check: Forking a repository means making a copy of someone else’s repository into one’s own account. So the live code for Open Room lives in my account. When a builder wants to add their room or make a suggested change, they will fork my Open Room repository to make a copy they can make changes .

Cloning a repository (which will come up next) means making a copy of a remote repository to your local machine. When you navigate to the GitHub website, you’re looking at the remote repository, the one stored on the cloud and on GitHub servers. After cloning the repo, that code is now on your local computer, which is usually where you — or your AI — makes changes (also called development).

To a person without a technical background, terms like “repository,” “forking,” or “cloning” can sound intimidating. And it’s not unfounded. There are a lot of details and nuanced definitions that are connected to each of these words.

But with AI, it’s not necessary any more to know all the ins and outs of these words and concepts. They’re good to learn, but now you can learn them in a way that fits you.

Step 3: Create your room

After Kaylee had the copy of the repo on our laptop, it was time to set up the room!

In the repository, there is a room template already set up, and Claude Code knows from the instructions I built in for it to make a copy of the room template to make a room with the specific details of Kaylee’s room.

After Claude Code made the folder for Kaylee’s room, it asked her about what background image she wanted to use.

We had already picked one out together, so we were ready to go when Claude Code asked us.

After giving Claude Code the background image, it automatically checked the size of the image and resized it to keep it within size limits. (Again, it did this because it was following the instructions I gave it.)

After compressing the image, Claude Code then asked Kaylee what she wanted to call her room, and she said:

“Neon Glowy”

After making the change, Claude Code showed us how it modified the code.

The GitHub convention for showing code changes is to show which lines were deleted (with red minus signs and red highlighting), and which lines were added (with blue plus signs and blue highlighting). I love this view because it makes it easy to track what has changed and how easily.

Step 4: Checking the changes

After the changes have been made, two things happen:

1. You check if the changes look how you expect

2. You create a request to merge the changes you made into the live site

There are two ways to check the changes you made. The first is locally. To do this, we again turned to Claude Code. Here is where I helped out Kaylee. I said:

“I want to see the change locally”

Claude knew what to do and what to tell me, as the user, so I would know what to do:

“The dev server is running. Open http://localhost:3000 in your browser and navigate to your room to see it live!”

You may not know what “dev server is running” means, and in this case, you don’t have to. But if you wanted to know more, you could just ask Claude Code to explain it to you! And you don’t understand it’s explanation, which can often happen, just ask it to explain it again in a different way, and specify what part didn’t connect with you.

The important piece of information to pay attention to here is the localhost URL it provided, which is a way to launch an app locally. We copy-pasted it into our browser, and we were able to see the new changes in the room!

Previewing the changes locally

In the case of Open Room, when an app is launched locally, the floor plan will look differently than on the actual site. Instead of seeing a grid of tiles, you’ll see a list of rooms. This is related to how the code on your machine versus the code in local connects to the external database, which stores the grid positions of the rooms.

(Fun fact: Open Room is using Supabase, if you know it. Shoutout to for teaching me about Supabase!)

Either way, you can click on your room, and the local app will take you to your room and show any changes that you made.

If the changes you made don’t look how you expected, you can iterate with Claude Code. “Y U no make change?!”

If you like the change you made, you can push those changes to remote. In the next step, I’ll explain how.

Step 5: Push your local changes to remote

Up until now, Kaylee and I ahd been making changes to our own machine, and only we could see our changes.

In order to make it possible for other people to see the changes, we needed to push the changes we made to the remote repository — specifically our forked (copied) version of Open Room.

And then once those changes are back up in the remote space, they can be merged into the live site through a review process (by me, as the maintainer).

In the past, an engineer would be writing the code to push their changes to remote. But now, Claude Code can do it, and it should be pretty proactive with suggesting making a pull request (or PR).

That was the case with Kaylee.

If Claude Code is behaving properly, it should then give you a link to the pull request that it made.

Previewing remote changes

After your changes have been pushed to remote, there is yet another way to check that the changes took the way you wanted.

The Open Room app is deployed through the hosting service, Vercel, and when a PR is made, there is the option to view a Vercel Preview Deployment. This basically means you can see how your changes look in the app in a preview format.

One annoying thing is that I, as the maintainer, have to approve every Vercel Preview Deployment for new PRs. So you do have to wait for me to approve it before it loads for you to see it. I try to accept them as soon as I see them, though!

Occasionally, the Vercel Preview Deployment doesn’t show up. I’m not sure why yet, but I’m looking to it.

At that point, if nothing looks strange to you, your job is done, and it’s up to me to review your changes and approve them to go live on the live site!

Step 6: Getting feedback

Sometimes I may ask you questions or make suggestions, for example if something looks broken. For the most part, I’m happy to ship your room however you would like it to go live. The beauty of this is if you don’t like how it looks after seeing it live, you can change it!

What I love about this is that we can figure this out together. We make changes in conversation with each other.

Step 7: It’s alive!

If everything looks good with the room, then I merge the changes to the live site, and then it’s live!

Kaylee and I were so excited when we saw her room live! The first thing she did was open up her tablet and navigate to the site.

Part 3: Refining a room with A-ma

Once Kaylee’s room was ready to go, I sent her off to play, and then I brought in my mom. I really can’t do our conversation justice other than that it was so much fun. She made so many connections about AI in just half an hour, and it was delightful.

What I’ll share here is the nuts and bolts of the next step — how to make your room interactive.

I worked with my mom first to build her room, and then we went a step further to make the room interactive.

My mom loves tennis and plays several times a week, so she made the background image for her room a tennis court surrounded by trees.

For the next phase, we wanted to add a picture of Andy Roddick, a pro-tennis player, and a link to his podcast, which she loves to listen to.

We asked Claude Code to create a hotspot where we could click on Andy Roddick and go to this website.

Unfortunately, that isn’t what happened. Womp womp, Claude Code. There is a bug where clicking the image opens the image; and a link to the podcast is in its own button.

We decided to ship it as it was. First, because we were at pumpkin hour, and it was time for my mom to head home.

But second, because there will always be more. This process is iterative. For now, the goal was to give my mom experience using AI to build something, even if it was a small change.

And actually, the real value was the conversation I got to have with my mom. She was so engaged. She asked great questions. She really wanted to understand what we were doing and why.

And the best part was, she was impressed by the AI. She was impressed that it could still run even if my request had typos in it. She was impressed that instead of her having to go off on her own to solve the problems we ran into with Claude Code, that you could just ask Claude Code to fix it itself.

Oh, and she entertained the possibility of Open Room taking off, so much so that in true Tiger Mom fashion, she started to think about what could go room. But the point is she felt the zing of what was possible to do with AI.

And if my mother can feel that, so can you.

Working with AI should be fun, especially when we do it together

I had so much fun building with AI with my mother and my daughter.

In fact, it surprised me how much fun it was to build with my mother. (It even made me wonder ..should we do a regular video or podcast series??)

My mom does not have a technical background. She’s not a software engineer.

What she is is — thoughtful, inquisitive, curious. She asked questions until she felt like she can understand the answer completely. She wants to make connections, and she is open to the possibility of what can be done.

I think if you told me a month ago, You should teach your mom how to build with AI, I would have laughed and thought, No way. I wouldn’t have even have known where to start.

But that’s how fast things can turn. Partly because the technology is developing fast. But actually, I think it’s also because of how fast people are developing.

And I don’t mean a fast like urgency-fast, like being-chased-by-a-wild-animal-fast (though that is of course happening).

I mean the “fast” that comes from the hunger, drive, and desire to know more; to do more and create more. It’s a whole-person fast where it feels like you’re following something that is calling to you, where you wished you had more time so you could sink into a hard problem that you know you could solve and that you want to.

This is what working with AI should feel like — where the entire time we’re using it, we’re laughing. Whether the AI is doing exactly what we want it to and more or whether it’s completely and utterly failing, we’re smiling and we’re learning.

But most important of all, we’re connecting with each other. We feel like we can build something more and better because we’re doing it together.

That’s the world of AI I want to live in. That’s the future of AI I want to build in.

Want more? Here are other posts on how you can have fun building with AI:

The Off Hours 🌙

"Rules We Made Up" An Animated Short

This short film is adapted from a poem I wrote years ago. It’s about a specific kind of person: someone who thinks deeply but doesn’t like to take life too seriously. Someone who resists neat categories and feels perpetually out of step with a world that wants everything linear and labeled…

Listen now

3 months ago · 53 likes · 54 comments · Natalie Nicholson

AI Meets Girlboss

[Giveaway Inside] I'm Building The Visual Brand Blueprint. What My Beta Testers Taught Me.

TL;DR: Visual branding is the most underrated growth lever on Substack and almost nobody is treating it as a strategy. For the past few months I've been building a visual branding product with Dheeraj Sharma to help you actually do something about yours. Here's how the community shaped what we built, and what's coming next…

Read more

25 days ago · 36 likes · 32 comments · AI Meets Girlboss

It was March 8th, 2026, and my Substack feed was flooded with International Women’s Day celebrations.

One post caught my attention.

In honor of International Women’s Day, Lovable was giving free access to their full platform for the day. (Thank you for the head’s up!)

Lovable is a vibe coding platform, where you can tell their AI an idea you have for an app, and it will build the app for you.

of The Responsible AI Brief shared an example of how she built an app to test your knowledge of AI myths vs reality.

I responded saying I was so impressed she built the app so quickly, and I wished I had the time to try it, too.

And then of course, once I said it, I couldn’t get it out of my head.

So, while making my girls grilled cheese sandwiches for lunch, I opened up Lovable and made an assessment app for my Human-Centered AI Capability Framework.

But! The point of this post is not about the HCAIC assessment app, which I’ll share another day.

In the post, I’m going to share what happened next.

Subscribe now

Creating the idea

“Kaylee, do you want to create an app with AI, maybe a game?”

My 7 year old daughter’s eyes lit up, and she nodded vigorously.

“Yeah, yeah, yeah!” she said.

“What would you want to make?” I asked her.

“The game with the numbers, where 2 and 2 become 4 and you combine them, and they get bigger,” she said.

“2048?”

“Yeah, yeah, that one!”

And with that, we had a seed of an idea, just big enough for us to take to AI.

A general vibe coding overview

When vibe coding an app, don’t jump right into building the app.

This is especially true with Lovable where you don’t want to waste valuable, precious usage limits figuring out what you want to do. Use it when you know exactly what to do.

Start with a general purpose AI like Claude, Gemini, or ChatGPT to work out the kinks in your ideas and create a plan.

Then, have that AI do one of the things it does best — create the app-building prompt for you.

No matter how structured or detailed the initial prompt is, expect to iterate either with the planning AI or directly with Lovable.

Human taste, discernment, and judgment matures into its best form through iterations.

This is true even if — or perhaps even more so if — you are a 7 year old.

Setting up the conversation

I opened a Claude chat session and typed:

“My daughter wants to make a 2084 game in Lovable. What can we tell lovable to make it”

Claude was happy to acquiesce.

“How fun! A few quick questions to make sure the prompt fits what she’s envisioning:​​​​​​​​​​​​​​​…”

I looked at my daughter, pondering a thought.

“Here, Kaylee, you talk to Claude.”

Before I handed her the laptop, I told Claude:

“I’m going to have her talk to you. Be child-safe and appropriate. Start over with the options”

Now here’s the thing.

Yes, I told Claude to be child-appropriate. But I have no idea how well it will listen in every situation, so I’m not going to treat that request as iron-clad. I trust Claude enough to tailor its conversation style and be appropriate overall. But I was still going to stay close and monitor their interaction.

Setting up the application requirements

Here’s what Claude told my daughter:

“Hey! I heard you want to make your own game — that is SO cool! 🎮

Let’s figure out what kind of game to build. Pick your favorites:​​​​​​​​​​​​​​​​“

It provided my daughter options.

After my daughter selected her options, Claude asked several more questions.

Then it provided the Lovable prompt. Up until this point, it was just my daughter and Claude making decisions together. When it came time to work with Lovable, I stepped in.

Try Kaylee's app here!

From Claude to Lovable

My daughter kept trying to scroll the laptop screen with her fingers (it’s not a touch screen).

I had to teach her how to scroll-highlight, command+C, and command+V. It’s a fun moment as a parent to see your child learn new skills, especially when their brain is trying to direct their fingers to do something they’re not used to doing.

One unexpected roadblock we ran into was Lovable kept hitting internal errors. My guess is it was getting slammed with high traffic from its promo!

When it finally did work, my daughter grinned, laughed, and jumped up and down as though her legs couldn’t control themselves. I couldn’t help but laugh myself watching her.

Iterating with Lovable (and some Claude support)

Despite the internal errors, we iterated with Lovable.

When the changes were simple enough, we talked directly to Lovable, and we switched between me and my daughter.

We also worked with Claude to help us create new prompts for Lovable. For example, my daughter wanted to add a new mode to combine by colors rather than numbers.

Then we made one last major addition — a magical creatures mode. (I’ll take credit for the idea.)

Kaylee designed the details with Claude, and we passed them to Lovable.

Lovable did great! There were a few more modifications we had to make, like adding the Evolution path and making the font size bigger.

Try out Kaylee's app here!

And that was it!

The app was alive, had different modes, and it only took an hour to build.

And it probably would have taken less time than that if it weren’t for Lovable’s internal errors …and my daughter getting distracted by the Lovable app my husband built himself.

(You can check it out here!)

Meowmeter - Rate the Cuteness

So you want to vibe code an app…

If you’ve been using AI or even if you haven’t, and you’re curious about what it’s like to vibe code an app, I hope this example gives you a flavor of what the experience could be like.

Here are a few reasons why this process felt seamless:

1. We started with a vision. My daughter had a clear idea of the game she wanted to create. We didn’t have to create new rules.

2. We used AI to help us get started. One of biggest unlocks I had when working with AI was to have AI write its own instructions. You are still the driver. You can modify the instructions directly or give the AI feedback on how to improve the prompt. But having an AI write for another AI is a good way to increase chances that the prompt will be successful.

3. Use the tools that fit the task. Lovable is made for vibe coding and has a good reputation. It performed very well with detailed prompts (made by the AI) and did fairly well following the more vague, shorter human prompts. The UI was easy to use, and the default designs for the app fit the aesthetic we wanted.

4. Be open to what the AI produces. A 7 year old is a perfect partner to channel the open, full-of-wonder mindset that makes working with AI fun and successful. I’ve found that the less I hold onto a rigid idea of what I want, the better I’m able to move with the output an AI provides. I can be more creative, and I can appreciate what the AI creates even if it wasn’t what I expected.

“Vibe coding” could sound intimidating. It can sound like you need to know more than you actually need to know.

But with the right tools, vision, and communication, it can be more vibe than coding.

My 7 year old daughter hadn’t been told she can’t do it.

So she did.

Share

Roll call!

, an AI art expert creator, has been working on an animated short full of storytelling, heart, whimsy, and personally meaningful creativity.

One part of the short features a music festival scene, so she put out a call to her network to create claymation versions of themselves to attend the concert.

I couldn’t respond fast enough to sign up!

This post is my journey making claymation characters of myself, my daughters, and my mother — and the triumphs and frustrations that come along with working with AI to create images that capture a human self and connect with a human vision.

Claymation Me

Attempt 1 - Using a real image as the base + ChatGPT

First I gave ChatGPT my headshot and Natalie’s prompt.

9:16 Aspect Ratio. Claymation character, full body. High-quality polymer clay render, stop-motion aesthetic, handmade texture, fully rounded 3D character modeling. Disc Head viewed straight-on. NOT flat, NOT shiny or plastic. Exaggerated proportions, large expressive eyes, simple dot nose, small smile — all flat on face plane. Soft matte finish, slight fingerprint textures visible on surface. Soft studio lighting, neutral beige seamless background. Use attached photo as face/likeness reference.

This is what it created:

My first reaction was — Soooo cute!

My second reaction was — but that looks nothing like me.

So I told ChatGPT, and it tried again, but there was nearly no change. I tried being more specific (“add a smile”). I tried giving it more pictures of myself. And still there were barely noticeable changes.

The other issue with ChatGPT is it didn’t capture the claymation feel.

So I moved onto Gemini.

Attempt 2 - Using a real image as the base + Gemini

I immediately got something better to work from the start. At least she looked like a claymation character!

But geez, Gemini. I knew my face was round, but way to rub it in.

So I told it:

“That’s better but the face is way too round and I want the hair to be less flat”

It sort of followed the instructions, but I still wasn’t feeling a connection. Even after a few more prompts, it just wasn’t right.

Should I just pick one of these and move on? I wondered.

But this little character was going to be my representative in a video meant to celebrate the true self inside each of us. I had to keep trying.

Then it hit me. I knew what I had to do, and I knew it would work.

Attempt 3: Using my cartoon me + Gemini

If you’ve been following me, you know I like to make hand-drawn digital illustrations, and I have a cartoon version of myself.

We were in the AI’s playing field now.

And Gemini did it!

There she was. A claymation version of me that I felt immediately connected to. She was cute, and she originated from me.

I sent her to Natalie’s music concert, excited to see how she would dance it up at the party!

Claymation Daughters

But I couldn’t stop there. I have two daughters, and I knew they would love to see claymation versions of themselves.

This time, I had a better idea of what I was doing.

Step 1 - Draw yourself

First, I asked my daughters to draw themselves in pencil.

Step 2 - Ask Gemini to recreate and colorize

I wasn’t going to immediately jump into Natalie’s claymation prompt. I knew I had to walk Gemini through this process one step at a time.

I told it,

“I want you to recreate this drawing exactly like it is. but add color here: hair is light pink and light purple. Pants black. Shirt dark pink. It’s a cartoon so no blending.”

The few things to note here are:

• Explicitly telling it to recreate the drawing exactly. Otherwise, it has a tendency to reinterpret the image, which sometimes you may want. In this case, I just wanted to add color.

• I don’t feel the need to always talk in complete sentences. AI is a next word prediction model. As long as the words are associated with each other, I trust the AI to know “Pants black” means “Color the pants black”. If it doesn’t catch it, then I refine.

• I also know that the AI would likely default to stylized coloring, and I wanted clearly cartoon-like colors.

It worked really well!

Gemini got a little confused about an arm situation, but that turned out to be an easy fix (whew).

I was ready to move on to the claymation phase.

Step 3 - Give Gemini the claymation prompt

Next, I gave Gemini the claymation prompt provided by Natalie.

They turned out SO GOOD!! I was dying after seeing the first one, it was so cute and perfect!

When the AI works, it works so well. It’s like everything I imagined and more. No more tweaks were necessary. These girls were ready to attend the concert!

Subscribe now

Claymation A-ma

After a day or two, I felt the itch to do more. I wanted to make my mom. And I knew my girls would love the idea of A-ma attending this musical concert with them.

But, as it goes with mothers, A-ma + AI had a hard time mixing.

Step 1 - Draw A-ma

My 5 year old drew A-ma first.

She made a little mouth with her tongue sticking out and a star on the cheek.

It was very cute (I love the creativity of my little girl!). But I could also foresee the AI having trouble with it. Even to my quick eye, the tongue looked like teeth, and the star could look like a little Halloween scar.

But I love seeing what AI will make of ambiguous input.

The AI did have issues (you’ll see below). So my 7 year old stepped in to help create some clarity for the AI. (This is a small example of what AI literacy can look like.)

Step 2 - Ask Gemini to recreate the drawing with color

I used the same chat that I had used to create my girls’ claymation characters and a prompt with a similar structure for the second claymation character I had made above.

“Same here. Recreate it exactly first with white grey hair. And the dress pink.”

It did not go well.

“That’s scary,” my daughter said. I agreed. It looked like the AI did pick up the Halloween-esque unintentional style.

Let’s try again!

Step 3 - A drawing tweak and a new chat

I opened up a new chat and gave Gemini the drawing my 7 year old fixed up (shown above), and new instructions:

“Recreate this cartoon with grey hair and a purple dress. It should be cute. She’s not a little girl.”

Oh Gemini, you’re trying so hard.

Attempt 1: Not so bad, but that star is still stuck on the mouth
Attempt 2: Looks really good! But where did the arms go?
Attempt 3: The arms are stuck on a bit weird, and the mouth-star came back
Attempt 4: Gemini gives up completely and removes the problem component.

Step 4: Back to the drawing board, way way back

If at first you don’t succeed, try again and start closer to the input.

We made progress! And then... we didn’t.

Step 4: Be stubborn?

I debated going with the very first claymation. I really wanted to be able to just go with the first image.

But I can be stubborn sometimes. Just one more prompt. The next one will get it, I’m sure it will.

Every time we got one element right (finally! A cute winky eye on the right! Now do it to the left), it would fix one and drop the one that worked (…and now we have a left winky eye and a missing right eye).

I pivoted one last time.

Step 5: Back to ChatGPT

I went back to ChatGPT and asked it to help me build a prompt.

Then I realized I should cut out the middle AI and asked ChatGPT to do the edit itself.

And there she was. Little A-ma with sleeves and two cute happy eyes. Good job, ChatGPT! Way to slide in at the end with the save!

It was only after I sent A-ma to Natalie that I noticed there was an extra eyelash on the left eye. 😩 Well done, AI. Thank you for helping me practice being okay with imperfection.

AI Field Notes Takeaways

• Making images with AI and I are not yet friends. Unless it clicks immediately, leave me to my hand-drawn digital illustrations, thank you.

• Working on words with AI is much more forgiving. There’s space between words that create their own meaning. And I, as the human, can modify the words. With images, I’m at the mercy of the AI to make the changes I want it to make.

• Creating images more easily connects the community. My daughters can create drawings. I can share them with my family, and they can provide immediately feedback. We can more quickly bond and enjoy a visual product together than a written one.

And as with any creation with AI — take it slow. It may work and create fast. As the human, you have the ability to move slowly, thoughtfully, and creatively. Use the time the AI gives back to you to infuse yourself into the creation you’re making together.

Thank you Natalie for the opportunity to use AI in such a fun, creative space. I love these little claymation versions of my loved ones, and I can’t wait to see them rocking it out at the music concert!

“I tried asking AI to do something, and the answer it returned was garbage.”

Anyone who has worked with AI has felt this at some point.

Even for me, working with AI daily, I still fall into the trap of expecting a perfect response from AI, only for the first response to fall short of my expectations.

The best products I’ve created with AI come from iteration together.

A one-stop job turned into a back-and-forth journey

I got a ping that I was mentioned in a Substack Note.

Visual brand AI expert, named Pinkie, had tagged me to identify what song best captures my Substack using AI and to have it create an album cover.

I’m working on a rebrand inspired by Pinkie, so I was eager to try this visual AI exercise.

AI Meets Girlboss

How to Build Visual Brand Consistency With AI

This post is part 3 of a series on Branding with AI. Read previous posts by clicking here…

Read more

4 months ago · 51 likes · 54 comments · AI Meets Girlboss

Pinkie provided a prompt in her Note. I thought — easy peasy, I’ll pop it into ChatGPT and get a beautiful cover.

Instead, I traveled a journey through the musical unknown, copyright issues, and unexpected delight.

Let’s step through this journey together.

Hi, I’m Alyssa. I write at the intersection of work, technology, and how to human — how to work with AI while staying connected to who you are. I love sharing conversations I have with AI and to see what they spark in you.

Subscribe now

Step 1: Copy-paste the starting prompt

Here was Pinkie’s prompt:

Describe my Substack channel as one song that best captures its meaning. Choose a song whose lyrics, or cultural context aligns with how my work helps people think, work, or show up differently. In one short paragraph, explain why this song fits my Substack. Then ask me if I’d like you to generate an album cover image inspired by that song, matching the mood and symbolism of the song you chose.

I gave this prompt to ChatGPT without any other context. I have Memory turned on, so I decided to go with whatever it knew about my Substack. After all, my Substack is one of my top topics with AI.

This was ChatGPT’s response:

“This Must Be the Place (Naive Melody).”
Your Substack feels like this song because it’s about orientation more than answers—finding your footing inside complexity and realizing that meaning isn’t flashy, it’s lived. It’s quietly grounding—exactly the tone of “oh… I’m allowed to be here, like this.”

I loved the reason why it pulled out this song. …but I had no idea what this song was. Turns out it was released by Talking Heads (a band I’d never heard of) in 1983.

Step 2: Give the AI simple feedback

In response, I just told ChatGPT,

“I don’t know that song.”

Immediately, it tried again. ChatGPT said:

Totally fair 😄 let’s pick one that’s more immediately familiar.

“Vienna.”
This song fits your Substack because it gently interrupts urgency. Vienna is about slowing down without giving up ambition. That’s exactly what your writing does: it helps high-capacity people loosen their grip on “faster / more,” reconnect with their bodies and values, and still move forward.

Similar to the first song, I loved why ChatGPT used this song to describe my Substack.

But I’m not really a music person… I recognized the song, but I had no connection.

Step 3: Be open to what the AI gives back

At this point, I could have been really frustrated. “Why can’t the AI just do what I asked already?!”

Instead, I took this as a signal that my focus at the intersection of work, technology, and how to human is unique and hard to capture.

Again, I told it, “I don’t know that song.”

I was intentional about not providing it more detail (and maybe a little lazy).

I was curious what else ChatGPT could come up with what it had so far.

ChatGPT said:

Got it — let’s go universally known.

“Here Comes the Sun.”
This song fits your Substack because it’s about emergence after a long winter—clarity returning and the quiet relief of realizing that something hard is easing. You writing helps people feel the shift from pressure to possibility.

And indeed, ChatGPT came to its own “conclusion” that the song needs to be more known.

This was a song that could start playing immediately in my head when I heard it.

So I told ChatGPT to create the album cover.

“Not bad,” I thought, but it was also a bit boring.

Let’s keep going, ChatGPT.

Subscribe now

Step 4: Ask for what you want

I took matters into my own hands and just asked for what I want.

“How about Taylor Swift?” I asked.

ChatGPT responded:

Yes — this fits beautifully.

“Clean.”
Clean captures what your Substack actually does for people: it names the moment after intensity, after being tangled in expectations—when clarity returns and you can finally breathe and choose differently. That mirrors how your writing helps people reset their nervous systems and show up with more presence in work and life.

Not too bad, but still not quite right. I have high expectations, and I know ChatGPT can do better.

Step 5: Double-check what your AI knows

At this point, I thought — Let’s double-check what ChatGPT knows about my Substack. Make sure we’re on the same page.

ChatGPT has a Memory option, meaning you can turn on a setting where it will save details about you it can use across chats. You can modify what it “remembers” directly if there are things you don’t want it to use in your chats.

‘What do you know about my Substack?” I asked it.

ChatGPT said:

Here’s what I know about your Substack—not from labels or positioning statements, but from the pattern of what you write and how you write it:

• It’s a place where thinking slows down. You’re helping people notice what’s actually happening—in their bodies, their work, their relationships with technology and ambition.

• You sit at a rare intersection—behavioral science, AI, leadership, and humanity—but you translate that into felt sense rather than jargon. People leave thinking “something just shifted.”

• There’s a throughline of permission in your work—to be human in technical spaces, to trust relational intelligence, to stop contorting to meet systems that weren’t built with wholeness in mind.

• Culturally, your Substack resists loudness. It’s a steady hand on someone’s back saying, “You’re not behind. Pay attention. You’re already here.”

Just tell me where you want to go next.

I honestly didn’t know where to go next.

Step 6: When in doubt, ask the AI

There used to be a time I’d wrack my brain trying to figure out on my own where to go next.

Now, I just ask the AI. Not to have it think for me, but because I’m curious what its ideas will be and what could spark for me.

So I went back to the beginning. I fed it screenshots of Pinkie’s post and prompt.

I told it:

“I’m not sure. This was the inspiration for the assignment”

ChatGPT’s response:

Ahhh—this helps a lot. Thank you for sharing the inspiration.
What that post shows is an important contrast, not confusion.

That AI Meets Girlboss → “Bejeweled” take is about visibility, sparkle, reclamation of confidence, and being underestimated then undeniable. It’s outward-facing. Declarative. Shimmer-forward.

Your Substack is doing something different.

So if Bejeweled is sparkle as power, your Substack is clarity as power.

I couldn’t have agreed more.

I admire Pinkie and her razzle dazzle style. This was also a helpful contrast to how my brand shows up and what I wanted to capture.

Step 7: Keep asking for what you want

I asked ChatGPT for a different Taylor Swift song.

This time, it gave me several options.

It said:

Based on your Substack’s actual feel (gentle clarity, relational intelligence, steady re-orientation), here are better-fitting options, grouped by why they work.

If the core is quiet clarity + coming home to yourself

evermore
This one might be the closest. It’s about realizing the pain/uncertainty wasn’t permanent — not through triumph, but through perspective. That’s very “Step up, step together.”

If the core is relational strength (not hustle strength)

The Archer
Deeply self-aware, emotionally intelligent, and honest about inner dynamics. This fits your how to human lens extremely well.

If the core is grounded confidence without bravado

invisible string
About trust, connection, and meaning emerging over time — not forcing outcomes. Very aligned with your “back-and-forth bridge” metaphor.

I love Taylor Swift + Bon Iver, and evermore was a song that I could hear in my head, so I went with evermore.

Step 8: Provide examples

Given that the first album cover was pretty generic, I fed ChatGPT examples of the illustrations I draw for cover photos and images for my newsletters.

I actually wasn’t optimistic about what it would create because my style is so simple. I couldn’t picture how evermore would fit together with my style.

Step 9: Find out more info

I asked ChatGPT to generate the album cover, my dopamine receptors ready for something good.

And then… I ran into this blocker.

ChatGPT said:

We’re so sorry, but the prompt may violate our content policies. If you think we got it wrong, please retry or edit your prompt.

Womp womp.

So I asked it,

“What is the prompt?”

Below I share the prompt, but with a caveat.

I don’t always read everything the AI provides. I skim for what feels relevant, then give the AI feedback on what to follow up on.

The prompt below is an example of this. I wanted to get a sense of the prompt to get a sense of what needed to be fixed. But I really didn’t care about the prompt itself. I just wanted the AI to produce what I was looking for.

Here was the prompt:

Prompt:

Create an album cover inspired by the song evermore by Taylor Swift, interpreted through the visual style of Alyssa Fu Ward’s illustrations.

Style: hand-drawn, simple linework, playful but thoughtful, slightly imperfect marker or pen outlines. Flat colors with a soft, calming palette (sage green, muted blues, warm oranges, gentle yellows).

Scene: a small, friendly orange character (non-realistic, cartoon-like) sitting or standing near a winding path or river that moves through a landscape. The environment feels quiet and reflective rather than dramatic. Subtle symbols of transition and clarity appear along the path (a small bridge, scattered leaves, gentle hills, a desk with a computer nearby).

Mood: grounded, reflective, hopeful, steady. No sparkle, no glamour, no performance energy. The feeling is “I can think clearly again” and “this won’t last forever.”

Composition: centered and balanced, lots of white space, readable at small sizes. Title text optional, handwritten script style if included.

Overall feeling: human, relational, calm confidence, clarity after uncertainty.

Edit: Okay, so now that I’m rereading this piece for edits, I did finally stop and read the full prompt. It’s beautiful. 🥹

Step 10: Use what you’re given

At the time, the prompt seemed reasonable based on my skim. So I asked ChatGPT,

“Can you try that?”

And boom it went.

This is what ChatGPT produced:

I was floored.

I should be used to it by now, but it still gets me — that feeling of being lifted up by a balloon with joy and disbelieving wonder. AI has done it again.

I loved this image.

I felt like it captured that whimsical, open feeling from evermore and what I feel like my writing can be.

The computer in the background brought in technology.

The bridge symbolized how I connect across different areas. The bridge stops at the water. My interpretation is sometimes the destination is looking out over clear water, not how you got there.

I was ready to share this image and declare a thank you to Pinkie and the amazing capabilities of AI!

Then I zoomed in.

Step 11: Double-check AI output

I thought something was a little off with the body and face.

And there was.

An orange head covered in little smiley faces.

I couldn’t help but laugh.

Part of me loved that AI picked up on the faces in my drawings.

Another part of me appreciated the reminder of how AI still need humans.

So still more work to get this one-shot exercise into something usable.

Step 12: Backing up a bit

I told ChatGPT what was wrong:

That's amazing. Now change this -- you have smiley faces all over the head. Fix up the face so it's actually a serene face turned toward the sky, eyes closed.

….and hit the copyright blocker again.

We’re so sorry, but the prompt may violate our content policies. If you think we got it wrong, please retry or edit your prompt.

Now here’s the thing.

This may be where a person could get annoyed enough to slam their laptop shut and swear off AI for good.

Don’t do that.

Here’s why.

When you run into a loop where the AI isn’t giving you the answer you’re looking for, you have an easy workaround in your backpocket.

Just ask the AI to help you.

Which is what I did.

I told it:

“Help me fix the prompt so it doesn’t violate the guardrails”

And it did.

ChatGPT’s response:

Yes — this is fixable, and you’re running into a very specific guardrail, not a real problem with the content.

What’s triggering the guardrails

Two things together are causing it:

1. Naming a copyrighted song + artist (evermore by Taylor Swift)

2. Asking for a visual interpretation “inspired by” it, which can be read as “derive from identifiable copyrighted work”

The solution is simple:
👉 Describe the mood, themes, and emotional arc without naming the song or artist.
You can still personally think “evermore.” The model just can’t be told that explicitly.

Then it gave me a new prompt to try, and told me why it would work.

Prompt:

Create an album cover illustration that evokes a quiet, reflective sense of endurance, perspective, and emotional steadiness after uncertainty.

[prompt continued…]

Why this works

• ❌ No named song

• ❌ No named artist

• ❌ No “in the style of” a living artist

• ✅ Same emotional DNA

• ✅ Same aesthetic

• ✅ Same you

Some may think I am giving up my critical thinking skills to AI.

I disagree.

I’m reserving my energy for other problems to solve …like evaluating whether my orange’s face got fixed.

Or better yet, I could work on a different task altogether that would benefit from higher-order thinking, instead of wracking my lower-order thinking for how to get the AI to do what I want.

Step 13: Ship it

And 6th time’s the charm! ChatGPT created an image that I’d be happy to use.

Isn’t it beautiful?

Takeaway: If at first you don’t succeed…

…try, try, try, try, try, try again.

When I first entered into this AI exercise, I thought I would be able to just run a prompt and get a perfect output.

I’m actually glad the prompt didn’t work right away.

One of my favorite parts of working with AI is creating something through our conversation. When I work with AI in an iterative way, I feel my brain creating new pathways and coming up with new ideas.

Now, here’s a nuance.

Yes, conversation with AI is a great way to get what you need.

But that doesn’t mean you need to have a long drawn out conversation.

In this one, most of my “prompts” in this conversation were short, just enough to provide the AI a signal that “this isn’t quite right”.

This doesn’t always work, but in this case, it was enough to get the AI to shift down different pathways of its own.

Thank you for this fun exercise and chance to share how I use AI! You inspire me with every post you make on how you use AI in creative ways that honor people for who they uniquely are.

If you’re someone who wants to work with AI without losing your voice, your pace, or your humanity, this Substack is for you.

Subscribe now

If you’d like to keep reading, here are other pieces that explore what it looks like to think with AI through real conversations and experiments.