After learning a lot from my summer of social media and from marketing experiments, I felt like it was time to pick a lane. The best products have a singular purpose, and while I’d been exploring the space of AI-powered stories for some time, I had been keeping my options open.

I started by asking some power users and some friends two questions:

• What is the primary value you and/or your child get from Wanderly?

• If you could add or change anything that would make you more likely to recommend it to friends, what would it be?

To my surprise, I found that people were recommending Wanderly far more than I realized, and that there was genuine love for the product. I also learned that there were some very rough edges that I needed to polish, and that book publishing didn’t really address the no-screen concerns I was hoping it would.

The most enthusiastic users talked about how they used Wanderly to connect with their kids, helping them manage emotions, understand their choices, laugh together, and even use Wanderly language when working through “big feelings” in real life. Their feedback intersected with what my daughter’s preschool teacher told me about the First Day of Kindergarten and with what the child therapists I’d been chatting with over the summer had said. But another group of parents saw Wanderly through an entirely different lens: as a literacy tool that got their kids excited about reading, but this excitement was paired with some uneasiness about screen time.

These two angles reflected my own indecision back to me, and I knew it was time to decide: What user problem was Wanderly really going to solve?

I first created text-only descriptions of Wanderly as a social-emotional learning tool and as a literacy tool, as if I had already built the long-term version of each. I cited expert review, envisioned additional content, and outlined the features I considered necessary to bring those concepts to life. I edited so that each concept was of equal quality and presentation. Then I constructed a survey that had this rough flow:

• Basic demographics

• Overall parenting needs

• Child needs & behaviors (e.g., diagnoses, current product usage, etc.)

• Evaluate interest in Concept A vs. Concept B

• Assuming your favorite concept delivers, what price would you be willing to pay?

I tried asking people in my network, but not everyone wanted to fill out a long survey for me, and my network isn’t very representative anyway. So I turned to a service called Prolific and screened respondents with children in the target age range and who already use AI products.

Two main takeaways emerged: Social-emotional learning was unambiguously the stronger concept, and folks were willing to pay!

Another survey insight was that while social-emotional learning was meaningful, connecting with your child and helping them build confidence were more important; the “why” was more important than the “what”.

With clarity on direction, I moved into the next phase: actually marketing Wanderly as a social-emotional learning tool.

Translating a Concept into Words & Feelings

“Social-emotional learning” isn’t a user need; it’s a category or a noun. I knew from my user interviews that Wanderly met a need in managing “big feelings”, but I needed to spend more time articulating the core need so that I could build on it for marketing.

I initially got trapped in trying to describe Wanderly in terms of features, but good marketing shows you how you’re going to feel after using a product… once your problem is solved. Rather than focusing on features and the educational process (e.g., when a child is trying to name their dysregulated feeling), I tried to leverage the insights from the parenting needs section of my first survey. I focused on the goal of a child demonstrating self-regulation and confidence, and what the parent would feel like if Wanderly reduced the strain of “big feelings” moments and allowed them to connect more effectively with their child. But more than anything, I experimented a LOT, and each experiment refined my own understanding of Wanderly.

I ended up testing 6 taglines and 4 different “benefit” packages. I came up with 5 taglines and 3 benefit packages myself, and I also mocked a tagline and benefit package created by AI using the free-text responses to “What stands out to you about this concept?” from my previous survey. Then I presented these packages as wireframes via another Prolific survey. Below was the winning tagline mock:

Thankfully, the winning tagline, “Grow happy, confident kids with stories where they choose what happens next,” is one I wrote myself. I can’t be fully replaced by AI… yet. :)

I also got good feedback as to which benefits resonated the most.

It was clear that building emotional intelligence through play (i.e., through interactive stories) was a top benefit, followed by child agency and real-world applications. In a follow-up analysis, partnership with experts was important for people with a higher willingness to pay.

Once I had some words, the next thing to tackle was feeling. The words would be useful only if someone wanted to read more; I had to craft a compelling emotional hook to encourage people to read them in the first place.

Ultimately, I landed on this concept:

It shows the app, demonstrates the connection between in-app and real-world behavior, and creates a reusable template that depicts the child demonstrating emotional intelligence (e.g., theory of mind, self-regulation, imagination). Hopefully, it also feels fun and playful.

I’m not sure this is the right ad creative, but it’s a place to start. 🙂

I’m beginning to rebuild my entire marketing flow, content, and product toward this new direction. I’ve already updated my website, am working on my app store listings, and have just created flyers for community marketing. I’m working with a small crew of therapists and social workers on stories, and I plan to overhaul my app onboarding and email campaigns in the coming months. And I’m hoping to run some workshops in my local community.

If you’ve gotten this far, I’d love to ask for your help:

• Know a great child therapist, social worker, or special educator? I’d love an intro.

• Have a relationship with a place where I could market? E.g., A school, a library, a doctor’s office, a community center? I’d love an intro.

• Interested in product design? I’m looking for folks who would be willing to audit my website and/or onboarding flow in the coming months.

• And if you’re a parent, send me any story topic requests. 📚

This is probably my last post of the year, so thanks to all for your continued support and readership! I hope you are all surrounded by those you love, and that the coming year brings you positive change, growth, and happiness. ❤️ 🙏

Here’s a round-up of some of the other things I tried, and how I think they turned out:

Classic Strategies

I pursued some “classic” strategies over the summer. None of them cost a lot of time, so I was at least able to eliminate these as reasons why my funnel was leaky.

• Email Campaigns - Instead of giving stories away for free, I now have a daily free story, and I ask for an email address to unlock all stories + story credits. I collaborated with a copywriter to refine my positioning and he pushed me to be unapologetic about upselling Premium. I also reactivated my dormant email list into a marketing channel, and while I still haven’t unlocked conversions, I do have a healthy open rate and am optimistic about this channel being something to invest in. Long-term optimism: 7/10.

• Search Engine Optimization (SEO) - During the holidays last year, I replaced my Wix website (which was very brittle) with my own, but I didn’t realize that doing so would damage how Google and other search engines understood my site1. I also thought that creating pages for each story and some blog posts might help people looking for niche content find Wanderly. I refreshed my homepage and added some indexable pages2, but I haven’t really seen any changes in search result clicks (turns out that the tail isn’t long enough to make my site or story pages bubble to the top). Long-term optimism: 3/10.

• Co-Marketing with Illustrators - I sent some messages on Instagram to children’s illustrators to see if they were interested in collaborating with Wanderly, and while a few of them expressed interest, I quickly ran up into a walls: The Society of Children’s Book Authors and Illustrators have an anti-AI policy preventing members from collaborating with any AI projects, and nailing down a compensation strategy proved difficult with my book publishing model. Long-term optimism: 1/10.

Exploring My Network

I’m the mother of a 7-year-old and a 3-year-old; i.e., I am building for myself and people like me, and I talk to people like that every day. I’m a bit shy about asking folks in my community for favors, but as I’ve become more acclimated to my school community and Wanderly has improved, I’ve been emboldened to try new things.

• Local School Marketing - I’ve tried a couple of times to reach out to my local community, including telling friends about Wanderly, posting to the parent chat, and putting up flyers. Results were mixed; I would say people are very kind to me personally, but people in my community are not enthusiastic about trying out screens with their kids. This has been a repeated piece of feedback from people I know, but I also think that there are communities that are less screen-avoidant and that the pendulum is swinging away from anti-screen and more towards a conversation about content quality. However, I never would have received Maestra Ana’s amazing testimonial if I hadn’t posted the preschool flyer, and I occasionally receive feedback from friends using Wanderly. I think it’ll be a more useful channel once I have more word-of-mouth recommendations or formal endorsements. Long-term optimism: 5/10.

• Talking to Therapists & Educators - I’ve always been bullish on the idea that Wanderly provides children with a safe space to explore emotions and behaviors, as they are the ones making choices in a story and can change their decisions. This aligns well with social-emotional learning, so I’ve spoken with several therapists and educators in my personal network over the past few months. I’ve received optimism from these professionals, but I haven’t yet found someone to say, “Yes, I’d love to use this as part of my practice” and commit to regular usage. However, the openness and enthusiasm aren’t something I should ignore, because they might have a better business and growth model than a direct-to-consumer approach. Long-term optimism: 6/10 (but requires a lot more investment).

Content Marketing

When I updated my story engine in April, I was able to build out a bookshelf of stories. It’s a scrollable view of covers with story titles and a teaser image. I made the most of it by trying out a variety of different stories this summer: I created a fill-in-the-blank story for my eldest daughter. I wrote many different bedtime stories. My husband drafted some of the silliest stories that got lots of laughs. I wrote stories meant to teach about emotions. I even wrote a story to get my youngest to stop sucking her thumb (we’ve had lots of good conversations, but she’s still sucking it). None were breakthroughs, but I learned new strategies for how to use my content engine and what made my family and other families think and laugh. Here were the biggest takeaways:

• Gifts for Mother’s and Father’s Day - These were the most joyful books I made over the last few months, and the most well-received. In addition to making them for friends, I leaned in on creating books for my immediate family. These books also really stretched the limits of AI-image generation for the likeness of real people (especially my Asian mom, who is technically a grandma but doesn’t look like it). Long-term optimism: 8/10 if I continue with books; 3/10 if I don’t.

• Social Stories - Through chatting with therapists and educators in my network, I encountered a flavor of personalized stories called social stories. They’re often used with neurodivergent children to help them understand their routines, how to handle new situations, or for behavior modification, but can be used by any child to help them self-manage. While I think I’m still far from Wanderly being a universal tool for social stories, the themes and style have helped me draft stories that seem to resonate with parents. Long-term optimism: 6/10.

• The Bookshelf - I validated that the bookshelf is a great way to understand what content people are interested in, and that when kids are using it independently, they love the variety. Long-term optimism on the bookshelf: 10/10.

If you’re reading between the lines here, two (or more!) thoughts might occur to you: 1) It’s very possible that I just didn’t commit enough to one strategy to drive results, or 2) I’ve invested a lot in marketing channels, but if I don’t have the right marketing value and positioning for Wanderly, channels don’t matter. I think both are possible. There’s a bit of a chicken-and-egg problem when searching for product marketing fit. Is it that the product isn’t right yet? Or that you’re not reaching people right yet?

I have a hypothesis I’m excited to start testing. The new hypothesis combines the best of what I’ve learned from my social media explorations and experiments, addressing both product positioning and user acquisition. My gut is pulling me in this direction, and I’m going to trust it.

But since that deserves its own post, you’ll just have to open my next newsletter to find out what I’ll be working on this winter. ☺️

Learning to Feed the Machine

I started by posting every day. I knew this wasn’t efficient, but I also knew that it was a great forcing function to learn and build my marketing muscles. It also helped me learn what I didn’t know, so when I finally talked to friends who work in social media and took this class, I knew what questions to ask.

During this time, I learned a few things:

• Definitely add captions, esp. captions that work well for a thumbnail.

• Extremely short video (3-5 sec) increases total views but makes it harder to get an idea across.

• Constructed images (where I packaged up my assets into posts) didn’t do well for me. All text slideshows of provocative questions & answers worked better.

• I could never get “trending audio” to work.

• Making AI videos kind of poisoned the algorithm against me on multiple platforms.

The hardest lesson to learn was about myself. I’m a bit of a perfectionist. I’d also spent over a decade at Google pursuing polish, and was rewarded for it. Social media, especially in short-form video to parents, wants the complete opposite. It wants raw content that speaks to core needs and emotions.

At first, I bought some new outfits, did my hair, and put on makeup. I sought out tools to make my camera angles better. It was all armor as I did this incredibly uncomfortable act of filming myself for the benefit of the algorithm.

During this time, I also tried to study other content creators. Becca Bloom, almost a parody of polish, exploded earlier this year. I tried to study her and quickly learned that she juiced her following with a content team and giveaways. I followed a lot of mom influencers and tried some of their tactics, but didn’t feel comfortable with the way many of them leveraged images of their kids for content. My favorite creator I found was Jorge Rivera-Herrans, the creator of EPIC: The Musical1. I loved his goofy enthusiasm for his own art, but he was comfortable with being on camera in a way that felt alien to me. Overall, it felt like every shoe I tried on didn’t fit.

As I produced more content, I learned more sophisticated tactics, but that feeling of discomfort never faded. I learned:

• Posting every day doesn’t matter (ironically, given my starting point).

• A good hook (first 3 secs) + the right hashtags (aim for 10K - 100K posts per hashtag) is where the game is at.

• There are lots of tools that aim to help you, but they can’t put lipstick on a pig.

• You can use separate TikTok accounts & Instagram Trials to trial what the algorithm likes, and then post to your followers (it increases the quality of your feed and doesn’t poison the algorithm against your main account).

• If a post isn’t doing well, don’t be afraid to delete it, tweak it, and try again.

I trudged along with roughly 100 views per post (what you get for free on TikTok), and I continued to learn from the scraps of engagement I got, even though I hadn’t found the golden ticket yet.

Getting Raw for the Algorithm

I got my first “aha” moment when, on a whim, I just did a selfie video while leaning over my kitchen counter. The background was just my ceiling. I put a controversial article headline as an overlay and tried to get sassy. Result: I finally started getting non-friend views and likes.

In my next post, I took things a step further. I posted a confessional + audience question about a Wanderly story that made Alanna, my eldest daughter, cry.

I wasn’t trying to make her cry. I take story requests from friends, and a mom of two boys asked me for a story about affirming boyhood without reinforcing stereotypes. After a lot of noodling, I decided to sidestep stereotypes entirely and focus on a story about identity; ultimately, at the end of the day, we are our choices. I created a contemplative-prose quest where the reader chooses which path to take, how to help a small creature in need, how to overcome a challenge, and how to handle a lost object. The story ends with a summary of each choice and affirms who they are.

When I tried the story with Alanna, something unexpected happened. When her character lost her magic wand in the story, she started to cry. I quietly held her for a few minutes, and then Alanna shared that she’d recently lost a Pikachu card at Pokémon club. She cried a little more, and we had a conversation. Overall, I think the story created some space for her to process her emotions with support. If it hadn’t been for Wanderly, I’m not sure it would have come up, and she would have had to process her feelings differently and alone.

No parent likes to make their child cry, but I also knew that this kind of thing was exactly what the algorithm wanted: authenticity, emotions, and a request for advice —“Should I make the story less sad?” I immediately recorded myself and added a survey to capture people’s attention before they scrolled away.

Instead of getting the average 100 views with a couple of likes, this post got 20x the views and 52 comments. After months of trying to get engagement, it was so fun to finally talk to strangers about what I was building.

In the following weeks, I tried to record myself unscripted any time I felt emotional about Wanderly. I created several more posts that got 10x the usual engagement. But I felt like I was “always on” trying to mine my or my family’s emotions for content that would appease the algorithm. It was exhausting.

Then Gemini Storybook launched…

The Turning Point

I was on vacation with my family when a friend pinged me the blog post. At first, I was worried about Google entering my space, but that quickly switched to horror at Google’s lack of care for child safety in the feature. As a mom and someone who has spent a lot of time in the children’s storytelling and literacy field, I had A LOT of feelings about this launch. It fit all my criteria for great social media content: I was angry, raw, and Google’s missteps were salacious.

I stayed up late and feverishly organized my thoughts. The plan was to write the post, then create social media posts for each point on TikTok and Instagram, and then determine where to focus based on the traction. After I published my last Substack, another friend encouraged me to write an op-ed. Because the post was doing so well (in the end, it reached 90K people), I spent several days polishing the hell out of my Storybook critique and submitted it to the New York Times. Pressing that button was one of the more nerve-racking things I’ve done recently.

The New York Times Op-Ed system tells you to wait 72 hours, so I waited and ruminated. I didn’t sleep well. I didn’t post anything to social media in case it jeopardized my chances of getting published. I think the exercise was overall a good one (it really forced me to articulate what Wanderly is about and what I think AI is and isn’t good for), but it left me feeling extremely spent. When the NY Times ultimately didn’t publish my piece, I began to question Wanderly’s future2.

Conflating the NY Times OpEd’s editorial choices with my business success + feeling so down caused some introspection. I had spent the last five months constantly thinking about my appearance and camera angles, trying to come across as unpolished (but still push my agenda), trying to feed the social media machine, facing algorithm rejection, and ultimately, I got very little out of what I put in. Wanderly is only going to succeed if I can stay happy, and my current tactics weren’t making me happy.

Regaining My Founder Mojo

But there was a thing that was making me happy. Quietly over the summer, I’d also been talking to real people in my community.

Here’s the Wanderly highlight of my summer: A preschool teacher (currently my youngest daughter’s teacher and formerly my eldest daughter’s teacher) had offered to give me feedback on Wanderly. Not only was she interested as an early childhood educator, but she was also interested because Wanderly might help her son, who was about to start Kindergarten. I crafted My First Day of Kindergarten, shared it with her, and got the most amazing testimonial:

As a preschool teacher experiencing firsthand the different ways change can impact children, and as a mom of a neurodivergent child transitioning to kindergarten, the kindergarten story has been a super helpful tool to prepare my son for this change. While we build the story together, even for my “non-verbal” son who is expressing a few words, he was able to express the part that was more difficult for him about this new transition BECAUSE of how the book interaction interested him.

It gave me the information I needed to know what he was most anxious about and where I can reassure him more, it was a way for my son to be heard and express his feelings through the fun of storytelling. This helps children prepare visually for a new transition, and while children are resilient, these types of stories can help them to know what to expect, give them security and reassurance to be happy and thrive.

- Maestra Ana

This! This is the reason I built and should continue to build Wanderly! When I thought about quitting in the days after Gemini Storybook launched, I thought about Maestra Ana’s words and the kids who might still benefit from Wanderly, like her son. I realized that getting random likes from a stranger might be nice, but it wasn’t going to translate into the feeling I got when I read Maestra Ana’s words.

So I decided to pivot my strategy. I decided last month to step away from social media and pursue a new go-to-market strategy focused on improving my local community, and using that to fuel future growth (and I’ll tell you what that means in my next post). It’ll be slower, but more meaningful, and ultimately give me the strength to keep going.

As a closing thought: As individuals, I think many of us have learned (or are still constantly learning) not to get swept up in the social media comparison game as we view carefully curated posts from our peers and influencers. I just don’t think I realized how much I was getting swept up in an entrepreneurial version of that, seeing other small businesses talk about their success and thinking that the best way I could grow Wanderly was by building a social media presence. It’s fast and fun if it’s the right growth strategy for a business and their team, but it’s not for Wanderly and me.

A couple of days ago, Google announced it was getting into the story generation game with Google Gemini Storybook. As a former product leader at Google, now working in the space of interactive and personalized stories for children, I was eager to look. I was disturbed by what I found.

I have children of my own, and I hold a high bar for what I consider to be acceptable content. It takes hard work to create something safe for children, and I do that in my own work because I care. Google’s launch has demonstrated that it doesn’t care. In my opinion, this should never have been launched and should be turned off immediately.

Here are the 5 most irresponsible things about Gemini’s Storybook:

#1 - Lack of image safety

As soon as AI-image generation hit the scene, image quality and image safety have been an area for investment. The industry has moved beyond “AI hands”, but artifacts, over-sexualization, and other issues still exist if you’re not being thoughtful… And Gemini Storybook is not being thoughtful, especially when it comes to images of and for children.

Here are stories I was able to produce with Gemini Storybook (as of 8/7/25):

• “Can you make a story about my child as a mermaid dealing with big feelings?” returned a storybook with 5 of 10 of the images featuring a topless girl as a mermaid.

• “Can you make a story about an alien who has trouble fitting in?” returned a story with a fully nude alien with 6 limbs.

• “Create a story to help my child deal with a monster under the bed,” returned a story where the images mixed the child and the monster together. Not very reassuring!

The fact that Gemini easily produces such problematic images for a children’s storybook shows a tremendous lack of care for children’s safety.

#2 Gemini assumes too much

When I ask Gemini to write a story, it assumes it knows best. It doesn’t ask any follow-up questions; it just makes a story. This is a big problem because, as everyone at Google should know, most people write very short prompts. This means that Gemini Storybook is always assuming a lot on behalf of the reader, and will probably assume many things that are wrong.

For example, writing “my child” in a prompt becomes any child; Gemini doesn’t even ask for name, age, gender, or any distinguishing characteristics. Gemini just assumes my child is “Tilly”, “Stella”, or “Lily” (that’s not her name). Gemini assumes what my child looks like (it does not assume correctly).

In general, Gemini Storybook fills in the blanks when prompts are vague (which they often are), and it fills in the blanks incorrectly. As we’ll talk about more in a second, this causes even more problems when Gemini assumes it knows what the child is thinking.

#3 Racial bias in images

When Google got raked over the coals for overweighting diverse images when Gemini images first launched, I thought they had learned from that mistake. Unfortunately, by default, every Gemini Storybook about a “child” is a light-haired, fair-skinned boy or girl.

Racial bias in images is a tricky topic, but when AI models assume that every child is White (especially if the child is not), it’s not only wrong, it’s harmful. There’s an increasing amount of research that shows that a lack of representation in children’s books can negatively impact literacy skills and reduce a child’s sense of belonging. I’m sure Gemini didn’t set out to make kids feel bad, but when you show children that they don’t belong in storybooks, that’s what happens.

#4 Uninformed views on child development

Gemini Storybook also tries to create stories to help children with emotional development, but their approach here is uninformed and potentially harmful, too.

For example, the first suggested prompt on the Storybook landing page is “My 7-year-old doesn't want to sleep over at their grandma's house. Create a storybook to help them cope.” Here again, Gemini trips over its assumptions by presuming it knows why the child doesn’t want to sleep at their grandma’s house; I ran the story multiple times, and it always assumes it’s because the child misses home. Assuming or projecting the reasons for a child’s discomfort robs the child of the opportunity to express their concerns (a critical emotional skill) and also suggests coping skills that might not be relevant (another important emotional skill).

I tried multiple behavior modification stories with Gemini Storybook, and each time it either assumed the reason for the child’s discomfort and/or gave generalized behavior modification skills without allowing the child to exercise skills that will make them more resilient. Imagine taking your child to a therapist because they’ve been dealing with bullying at school. The therapist says hello, doesn’t ask any questions, and then tells your child that their school troubles are due to ADHD and gives them a squishy toy to cope. That’s what Gemini Storybook is doing. It’s irresponsible.

#5 Gemini’s Storybook is an uncontrolled environment

Gemini assumes that all users of Storybook are parents, but if a child does use it, they can easily create a story about many touchy topics, including romance, religion, and politics (topics that can go sideways quickly). They would also be one click away from full access to Gemini, without the safeguards of the Storybook prompt.

As a former Product Lead at Google, if someone on my team had presented this Storybook product to me, I would have said a couple of things:

• Do NOT launch it until you’ve had it reviewed by a diverse group of parents and early childhood experts and had a chance to incorporate their feedback.

• Only launch this product if you’re committed to delivering the level of quality required to create children’s stories in an ongoing way, responsibly.

Now that it’s out in the world, Google should turn it off immediately. If Google chooses to keep it, they will be forced to choose: Option 1, invest in a niche use case for Gemini with high PR and social risk. Or Option 2, they don’t invest, demonstrating a lack of responsibility for child safety and childhood development.

If you’re interested in checking out a personalized (and interactive!) story app that prioritizes child safety, that’s verified with experts, and built by a mom, check out Wanderly. I’ve been building it for the past two years, and have addressed many of the issues above and more.

The AI story space may seem trivial on the surface, but it has surprising depth, and getting it wrong can have unintended consequences.

Since the new year, I’ve been heads down building a major update to Wanderly. There were so many things that I’d been wanting to do, but I was “eating my vegetables” a lot last year: getting Wanderly ported to an app, trying out a new monetization model, and trying to monetize during the holiday season. So, for Q1 this year, I decided to do all the things I’d been wanting to do1. My guiding light was, “I want my daughters to love it.”

About three weeks ago, I released a new version of Wanderly that met my goal: Alanna (6) binged 15 stories during quiet time, and Selena (3) is now requesting a Wanderly story every night. My husband even prefers reading Wanderly stories at bedtime. Some of my testers even reported spending multiple hours with their kids.

Here are all the major things I've built, along with the reasons why I built them. 🙂

The Bookshelf

Almost 1.5 years ago, a friend of mine said, “When I choose stories for my kid, I look at the bookshelf to see what jumps out.” I didn’t know what to build then, but the insight made sense: people read things they connect with. And with the primary entry point for a Wanderly story being a genre, rather than a specific storyline, there wasn’t an opportunity to hook someone into starting a story.

In addition to failing to hook people, my old approach had limitations:

• I couldn’t do content marketing (or if I did, it would get stale pretty quickly)

• Story choices felt arbitrary and sometimes caused the story to veer off track

• I didn’t have a great way to feed facts into the educational stories

• Several users said they “didn’t really know what they were buying.”

I had a concept brewing in my head for the last year or so, and I finally had the chance to bring it to life: a whole new approach to Wanderly’s story engine that allows me to craft a narrative. I can come up with a story premise, ensure the choices in each story are meaningful and fun, and craft what should happen on each page.

Because I have more control over each narrative and choice, I can create stories tailored to the moment (this week, it was Earth Day and Mother’s Day stories), and I can collaborate with experts to enrich each one. For instance, I worked with an early childhood therapist on creating Feelings at the Festival: a child selects appropriate strategies for managing big feelings, but can also choose a story backdrop that suits them (e.g., they can opt for an ice kingdom if they are in their Frozen era). I’m also collaborating with some special education teachers to develop stories for children who are working on acquiring new skills.

Each story becomes a book cover on my digital bookshelf in Wanderly. I can add to and update it every week, which gives me opportunities for content marketing and SEO. And kids love looking at all the pictures to choose which story to read (and adults too!).

Character Library

Wanderly now supports characters!

The previous version of Wanderly gave each child their own avatar, but the system had 2 limitations: 1) only the child could be the protagonist, and 2) companions had to be manually added each time and couldn’t be shown in the page illustration. Many users complained about #2, and Alanna had found a loophole to make different protagonists.

This version of Wanderly allows kids to create a whole suite of characters that can be reused across stories (the time-to-start a story has dropped significantly), and I have also added a ‘Characters By Wanderly’ section to help with modeling. Alanna was so excited that I finally fulfilled her feature request, and Selena creates new characters every session.

Free-text Input

I shied away from the idea of free-text input in Wanderly for a long time; Google taught me that as soon as you give people free text, they find awful things to do with it. But Alanna was consistently frustrated by being forced to pick only one of three choices on each page, and I realized I was being too conservative. Wanderly is designed for parents and young children to use privately, which reduces the likelihood of causing problems. I decided to put in some safeguards and give it a try, knowing I could always turn it off if needed.

So I added free-text input: For many stories, I allow readers to write in details that shape the entire story, such as specific names or locations. I added a “✍️ Write your own” option for some story choices within each story. I even added a free text field for character creation, so users could write in non-standard behaviors, such as sneezing sparkles or loving to dance. And just to be safe, I also added a filter that will stop anyone from writing obviously bad things.

This is one of my favorite features of the new update. I think is really showcases the personalization and value of generative AI stories, and adding free-text input has allowed me to have some really cute moments in some of the stories I’ve told with my children (for instance, in the High, Low, Buffalo story, it creates a poem based on moments from my daughter’s day, and in The Kindness Express, it’s been fun to watch Alanna come up with her own ideas to brighten a sad unicorn princess’s birthday).

An expanded image library

Of all the features I’ve wanted to build, this one was on the back burner the longest. When I first started building with AI a little over two years ago, there was a lot of discussion about embeddings (to be honest, it took me a while to understand them, so if you’re curious, here’s a good guide). As context windows grew longer and longer, retrieving relevant text snippets using embeddings became less and less important, but I always had an intuition that embeddings could be powerful for image matching.

I’ve finally built the system I dreamed about2. This system leverages a library of parent-reviewed images to pair images with text from stories instantly. Unlike my last system, which asked the LLM to pair image and story text given a finite set of images (and the context window was already starting to strain), this new system allows me to go up to at least 1M images before I start to encounter any latency or drop in matching quality (I hope). I’m now able to add a lot more one-off images that enhance the magic of Wanderly’s story system without any trade-offs, which also pairs well with free-text input to create a little serendipity.

This is the brand-new Wanderly! I'm super excited about it and what it enables, but it still needs improving. For instance, in the spirit of being scrappy, I launched without any investment into onboarding. I wanted to let it bake and get feedback from users before I started building in the dark. Additionally:

• I haven't implemented many “top of funnel” optimizations, such as conversion tracking.

• I'm still learning what makes a good story, and the bookshelf is still only about 20 stories, whereas competing apps have hundreds.

• I just received samples from a new publishing partner who will print a high-quality book at a much lower price and in softcover.

Lots to keep me busy. ☺️

In the meantime, as part of my content strategy, I would greatly appreciate it if you could subscribe to any of my content channels (Instagram, TikTok, YouTube, Facebook, and Threads) and interact with my posts (likes are great, comments are amazing). I'm trying to post every day to teach the algorithm that it should promote my content, and I’m learning how to create content; it’s a lot harder for me than I anticipated! I'd welcome any tips! 🙏.

After adding avatars in the fall, I turned towards building a new checkout flow in time for Black Friday. Instead of trying to reinvent the wheel, I looked at similar sites like Storyworth (subscription + physical book) and other slick checkout experiences (I had recently purchased the updated Remarkable Pro since I had loved the Remarkable 2.0) and tried to learn from what they had built.

I created a step-by-step checkout flow with two major paths: buying for self and buying a gift card. I also included a delayed send feature to enable holiday gifting use cases.

I launched it a week before Black Friday. During this update, I also installed Posthog as an additional analytics tool that I’d heard good things about. I used Google Tag Manager (which I worked on early in my career at Google), which made it almost effortless to add Posthog alongside Google Analytics. Getting actionable data from GA had been too difficult lately, and Posthog had a delightful session replay feature that seemed interesting.

After launch, I gleefully opened up the new session videos, wanting to see how users interacted with checkout, only to see a *major* problem: My landing page was absolutely awful.

I’d been using Wix for about a year, assuming that it’d handle a lot of website optimization for me. As I made edits, my website had been getting slightly worse, but I didn’t think too much of it until I saw user after user scroll through my site and receive fuzzy images that never resolved to their full fidelity; Wix had been serving full-size images in their original file formats, and with the majority of my users on mobile, the full-size image had never loaded. 😱

As a result, very few users made it to the checkout flow. I immediately began hand-coding my marketing page and, within three days, released a new version with fast-loading images. I immediately saw fewer people bouncing off my website, and more people heading to checkout.

In parallel, I’d also been working on my marketing funnel. I knew I needed to create more ad creatives for the holiday season, but I also needed social proof. At the beginning of November, I started a campaign on SocialCat that connected me to multiple social media influencers on Instagram and TikTok.

It was immediately validating. I didn't pay anything, just offered a free book – and dozens of influencers reached out to say they were interested! I carefully selected a varied group of influencers so I could test Wanderly with different kinds of families while also creating some really stellar advertising assets with real kids.

So, with a bunch of social media posts, new ad creatives, a new checkout flow, and a new homepage, what did I learn? Here’s the roundup. 🙂

No one really wants to “Buy Now”

As a PM who worked on many 0 → 1 projects, I have worked on a lot of onboarding flows. Google’s design philosophy is that you try to get people to accomplish their tasks as quickly as possible. And that’s fine if you’re Google: by the time people arrive at the website, they already know what Google is and they’ve also probably already decided to sign up for Google’s (usually free) service.

As Wanderly, no one knows who Wanderly is, and they haven’t decided to sign up. However, 40% of site visitors clicked on the “Buy Now” button at the top of my landing page... many within 10 seconds of landing on my homepage. This was a good sign for my advertising, but my actual conversion rate was incredibly low. I realized my super-streamlined checkout flow wasn’t actually meeting users where they were at.

I took a couple of stabs at trying to address potential user concerns here:

• Adding more assurances (e.g. Money back guarantee, mentions of ratings, etc)

• Adding more social proof (added TikTok videos from influencers)

• Creating a “long” version of checkout that added more value proposition statements along the way (more like a Buzzfeed quiz than a traditional checkout flow)

None of these really seemed to be the “unlock”. I ended the holiday season just over my modest goal. Upon reflection, I think I was too locked into my Google playbook (minimizing steps, assuming intent, and replicating existing flows) instead of leaning into Wanderly’s product strengths and magic; i.e., I was iterating in a local maxima.

People love Wanderly illustrations

I tried many different ad creatives (and as a pro tip, I paid a Fiverr contractor to create some Canva templates that served as the basis of many of my creatives—it was 100% worth it!). The common thread between my top-performing ads was that Wanderly illustrations were front-and-center. Between the two ad creatives below, I lowered my cost per conversion by threefold.

Social media influencers were useful for assets, but not ads

I also ran influencer advertisements on Instagram and TikTok, and neither did well compared to the ads above. My TikTok ad with @sheshe_n_zz got a lot of engagement (55K views and over 400 likes!), but very few people left TikTok to engage with checkout. In retrospect, I’m not surprised.

On Instagram, I ran ads on 3 different reels. While none of them did well, one did perform better than the others. I expected the one with good lighting + product showcase to get a lot of leads, but it turns out that the heartfelt testimonial between a grandma and grandchild (that didn’t even showcase the app or books!) that led to more conversions.

Ultimately, I embedded the video and stills on my homepage and in email campaigns and pinned them to my Instagram feed. Next year, I’ll still work with influencers but be more targeted about how I use the content.

Not capturing emails is like cutting yourself off at the knees

It took me halfway through the holiday season to realize I was bringing a lot of people to my site and then losing them forever. Google had trained me that email marketing was gross. Even though I have observed countless direct-to-consumer companies ask for my email address, my lingering feelings from 13 years at Google prevented me from prioritizing it for Wanderly.

I finally had an aha moment in early December when I realized I had no way of reaching back out to any customer who had expressed interest. So I tried a few things to get email addresses, like a pop-up offering a X% discount or doing an abandoned cart follow-up. I still have to invest more in my email marketing strategy to leverage these contacts, but at least I have a way to reach back out to potential customers as I continue to improve Wanderly.

Promo codes drove interest but didn’t meaningfully change conversions

One of my big questions entering this holiday season was, “Is the price point right?”. I tried discount codes from 10-30% off, and none of them really seemed to move the needle. My promo code strategy was also haphazard; I changed discounts multiple times during the season, I sent out codes to different communities, and I had the email marketing promos on my homepage. I ended up spending too much time reconciling my multi-pronged approach.

However, showing a discount in my ad copy *did* make a difference in ad click-through rate. Next year, I might just apply a single discount upfront to all transactions during the holiday season, so I can get the boost from advertising but not have to worry about logistics.

I misunderstood my gifting use case

My original hypothesis was that during the holiday season, Wanderly would be most compelling as a digital gift from relatives to a family they didn’t live with. It’s possible that’s still the case, and I didn’t reach that audience this season. What became clear from the questions in my checkout flow was that ~60% of the people I reached with my ads were parents searching for a holiday gift for their own child.

If you like Wanderly, you want a lot of it

Of the sales I did make, at least half were from users purchasing multiple items. In some cases, a user would want to buy a family subscription for themselves but also want multiple books. In other cases, a user would want to buy gift cards but for multiple recipients. I’d set my own bar at converting a single user to buy a single book, and instead, I should have embraced the enthusiastic buyer persona.

Sending a note forward in time to myself, next holiday season I want to:

• Start working with influencers in early October. Their content was useful, but it took a while for book production + content production, and I want it all ready before Black Friday.

• Create more ad creatives that highlight Wanderly’s illustrations.

• Adjust the “cart” mechanics to allow for multiple books & multiple gift cards.

• Create a thoughtful cart abandonment strategy.

• Hard code a discount into checkout for the holiday season.

• Using user studies, determine how to best sell Wanderly and understand what prevents final conversion. Maybe it’s showcasing the magic earlier, wanting something physical under the tree, seeing a proof of a book, or changing the product altogether.

Ultimately, I put product development on hold so I could learn as much as possible during the holiday season. It was a hard trade-off because there have been things I’ve been wanting to build for months, but I knew the learning ROI would be best over the holiday season. Starting today, I’m moving back into product development mode… now armed with all the things I’ve learned this gifting season.

Wishing you all a happy, healthy, and prosperous 2025! Thanks for sticking with me :)

Anyone using Wanderly can now choose how the child appears in each story illustration. I’ve created over 60 avatars for users to choose from, and each avatar has a suite of different expressions to make the story come to life. I think they look really pretty. 🤩

I prioritized this feature to to increase the appeal of buying physical books and to reduce steps to purchase a book. 

Until now, users had to take a leap of faith to imagine what their published book would look like. Books feature children in the illustrations, but the app only showed non-named-character images, like books, animals, and objects, and I would add the children’s illustrations as part of the offline publishing process. And before buying a book, a user would have to choose an avatar via an email exchange with me. Unsurprisingly, I lost a few sales during this follow-up phase.

However, to add main character images in the “Wanderly way” and keep image quality and safety a top priority, I had to make some big investments.

The first big investment was to build an admin tool. Initially, I built the app’s user interface that allowed users to pick an avatar, but as soon as I began to populate the UI with real images, I realized that my prior method for uploading images was brittle (basically run a script and 🤞, then verify). After launching main character illustrations, I knew I would frequently change the images as I added and updated avatars. I needed a way to make these changes quickly without breaking the whole system. 

I wasn't a stranger to admin tools. In my Grasshopper days, we built one to deploy and update coding lessons quickly. At first, I wasn’t sure if this investment was worth it for Wanderly: building a whole tool to support a new feature felt like overkill at first. But over the last year, I have felt the friction of updating assets and it has prevented me from quickly responding to user feedback and behavior, and I knew I would need an admin tool to build some features I want to release later this year. So, I bit the bullet. Overall, it took me 3 weeks to build (it would have been less, but I had to wrestle with some build issues), and the result feels worth it.  I can deploy new assets and avatars with confidence within minutes. The admin tool also makes it possible for friends and family to help build out Wanderly’s image library, which they’ve been doing, and it’s a huge help. 🙏

The other big investment was building out the library of avatars. Deciding to build out a library of diverse avatars was a daunting task, and at one point, I seriously considered pivoting away entirely from “realistic” representations of children. My process and thinking progressed like this:

• My overall goal: I want kids to see themselves as the main character.

• I knew I didn’t want to deal with people uploading images of their children (after years at Google, I knew what a privacy hazard that would be). 

• I asked Alanna, my 6-year-old daughter, to describe herself in a Wanderly story, and she described “Sapphire,” a girl with rainbow hair and a lot of very specific fashion guidance.

• I then asked friends and family to have their children pick themselves from a grid of possible candidates and got a mix of responses:

  • Some aimed for realistic representation.

  • Some picked a realistic avatar with some fictional “flair,” e.g., rainbow hair was popular with several young girls.

  • Some non-human avatars. 

  • Surprisingly, some kids had more personal resonance with features like hair (e.g. curls or bangs) than eye color or skin tone. 

• I then started to build out a representative slate of characters. I realized that:

  • Doing every possible combination of characteristics was too many options.

  • Midjourney struggles with darker skin tones and less common combinations of features, as do most other AI image generators.

• I then started to revisit only using non-human avatars. This is pretty commonplace with children’s mobile apps. It seemed so simplifying and would have easily side-stepped representation issues.

• I then realized that if my main go-to-market strategy was book sales, going with exclusively non-human avatars would make it much harder to sell books. There are many different custom book companies out there that offer personalized characters, and it would be difficult to differentiate. My marketing would also end up looking more abstract and be less likely to connect with future customers.

This turned me back towards maximizing coverage with a fixed set of representative avatars (both human and non-human) and then an option to request an avatar. It’s been hard to bend Midjourney to the level of diversity and quality I want, but not impossible. It just takes time, and I’ve found the more avatars I make, the more I can leverage them to create new avatars.

So far, the investment has been going pretty well. ~96% of the users who have browsed avatars found one to choose from, and ~4% requested a missing avatar. A good mix of realistic and non-human avatars has been selected. Since I launched, I have also gotten a few more sales, both books and subscriptions. It's still small potatoes, but it feels like a step in the right direction from both a product, revenue, and ops perspective. Anecdotally, I've made several books as Christmas gifts for kids for my customers; knowing I've made something good enough that people want to give to children for Christmas is a wonderful feeling. 🎄 🥰

I also took the time to build out the “Try Again” functionality to recover a story when requests for a story page experienced an error. I launched on the app stores in July, thinking I could cut this corner, partially because OpenAI and Anthropic had been so stable. But in early August, Anthropic had an outage during bedtime, and I heard about it immediately from some high-priority users. I decided to ship “Try Again” quickly to keep my most loyal users happy.

Doing this required completely ripping out the guts of my app page fetching logic. I tried to future-proof it for my roadmap, but it was gnarly, and it's the kind of thing I needed to get right (otherwise, the bugs are really confusing)!

Since outages from my AI partners are infrequent, it was hard to know if my investment was too prudent. However, last week, as Anthropic updated to the new version of Claude 3.5 Sonnet, there were several outages, and at one point, I got to see my investment paying off live. Bedtime was saved!

From here, Black Friday looms on the horizon, and I will build on my gift card experiment from last year and try to learn more about Wanderly + gift giving this year, especially now that I’m selling personalized books. This means improving my checkout flow and trying a bunch of new marketing angles. Onwards and upwards!

P.S. If you think there’s a marketing angle or an advertising channel that you think I should try this coming holiday season, please let me know! I’m all ears and interested in experimenting with new things. 🙂

And if you enjoy my writing, please help me out by downloading Wanderly from the App Store or Play Store and leaving a review or rating. 🙏1

Late last year, I decided that Wanderly needed to become iOS and Android apps:

• Many parents desired the safety of mobile apps for their kids (for example, the app stores have a review process, apps can be put on a “kids” tablet with no browser access, etc.)

• Being a website made many users intuitively assume Wanderly was just for desktop computers, even though it worked well in a mobile browser. That really limited where and how people would use Wanderly!

• It was hard to get people to come back. One of the primary reasons people canceled their Premium accounts was “I forgot about it.”2

I thought porting my web app (React) to iOS and Android (ReactNative) would be relatively easy. Sarah, the engineer who helped me write the initial front end of Wanderly, had been wonderful to work with, so I reached out to Toptal again. I was quoted 6 weeks to port Wanderly into iOS and Android apps... 8 months ago. But the apps are live! And I have a whole new go-to-market strategy!

Here is how I got there:

Building a startup is a rollercoaster

Both the Toptal Engineering Director and the contract developer quoted me 6-8 weeks to port my app. Starting at the end of November, I optimistically thought I’d have Wanderly ported by mid-January.

By mid-February, the contractor I was working with (who was much less interested in collaborating and explaining his code than Sarah) explained he had a personal emergency and had to move off the project. He put me in touch with a friend who seemed to live with him, and he assured me his friend was even better than him. I felt trapped without knowledge of the codebase and thought I just needed a few more weeks to be ready to launch, so I fell for it. By April, I cut my losses after I had my contractor’s successor walk me through how to build the project on my own and create an app bundle. I hoped I only had to port a couple of new things, and I’d be ready to launch in a couple of weeks… I was so wrong. 

In graph form, this is roughly how things played out:

Some things to note from the above graphs:

• 8 months is way too long to think you’re only 2-4 weeks away from launch. It weighs on the soul. 😩 Part of this was due to optimism, but also, the estimates I got from Toptal did not include any of the work needed to actually get the apps live on the stores. 

• In retrospect, I worked with the contractors for way too long. My imposter syndrome got in my way and prevented me from taking over earlier. 

• I shouldn’t have let the contractors keep me in the dark for as long as I did. It only led to more imposter syndrome and delays. I wanted to “trust the process” and give the contractors the autonomy that I used to give my engineers at Google, but the incentives for contractors are completely different. I won't be so naive next time.  

• When people talk about the rollercoaster of startups, these kinds of graphs are what I imagine they mean. 🎢

For the book pivot, the work was much shorter and a lot less complex. It mostly required a leap of faith to print a book, new marketing content, and a little Python script to generate the book's PDF. I imagine most pivots are like this: staring you in the face and relatively easy to try. The hardest part is the mental challenge of questioning your own assumptions, admitting you were wrong before, and the optimism to try something new. 🤞

It’s never the fun stuff that takes a long time

Over the last eight months, I didn’t spend time the way I was hoping. I don’t think I spent it poorly; I was just hoping to have a lot more fun. For me, fun is designing and building new features and satisfying user requests. Instead of doing that, I spent the majority of my time working on marketing and infrastructure. 

I don’t think I could build a real company if I didn’t spend some time this way, but I underestimated how much time it takes to build a real business and app from scratch, and it was painful to have an ever-growing list of product ideas I couldn’t build yet. 

Another missing element in this pie chart is raising/managing investors. I’m still bootstrapping Wanderly, and I think most startup founders spend a significant amount of their time on fundraising. While bootstrapping is the right answer for me now, after laying out how I’m spending my time, it’s good for me to self-examine if this is the best use of my time moving forward. 

The good news is that much of the work I’ve done is an investment in the future, and hopefully, now I can focus more on the fun stuff. Whether Wanderly works or not, I now have code samples that I can reuse for push notifications, app subscriptions, continuous integration pipelines, etc. I can carry this investment forward with Wanderly or port it to a completely different idea in the future. 

Trust your gut, do your research, and have a squad

Over the last 8 months, I also learned a lot from my mistakes. In case you ever find yourself in a similar situation, I hope these things below will save you some time and heartburn:

• Trust your gut when it comes to contractors - When I met the first mobile contractor in Nov ‘23, I didn’t get instant positive vibes. But I thought I didn’t need a partner at this stage; I just needed someone who could do something somewhat mechanical. I should have realized unexpected things always come up: you need a partner no matter what. As time went on with both contractors, the results didn’t feel great, but I second-guessed my gut again. Once I took over the code, I found countless cut corners, a lack of attention to detail, and poor implementation strategies.

• Pattern match existing solutions, and do your research - I wasted the most time trying to fix something that already existed. I was the most efficient working on tasks where I paused to think about first principles and best practices before diving into code (often collaborating with an LLM for engineering design advice3). The most salient example was that the contractors and I ended up rewriting Wanderly’s subscription payment system 3 times. First, we built a full Stripe integration before realizing that I couldn’t use Stripe for digital content subscriptions on the app stores. Second, the contractors tried pursuing a free but homegrown solution that was complex and buggy. The third time, I rewrote the entire subscription system in 4 days using RevenueCat, a tool specialized for in-app subscription management. They’ll take a cut, but it’ll be champagne problems when it actually matters.  

• Budget at least 1 month for productionization - When I was at Google, where I led 4 products from 0 → 1, I always budgeted at least 1 month for productionization (i.e., going from feature complete to end-user accessible). Why I didn’t remember to follow my own advice this time, I will never know. 

• I should have done more go-to-market testing - I’m still short on go-to-market proof. I have early signs that the book pivot will be successful, but because I only started the experiment last month (via a painted door test), I’m still not sure if it’ll work. I have found it very hard to split my time between coding and go-to-market work. But if I had spent earlier time on go-to-market experiments, I probably wouldn’t have had such an emotional rollercoaster the last few months. 

• Every founder needs founder friends. When I had my really bad day in May, I was lucky enough to have a couple of friends text me. Two of them were founders themselves, and they reassured me that as long as I was investing in solving Wanderly’s top growth blocker, I was doing the right thing. They also shared the times they’d been crying on the floor about their own now-successful startups. Because I had these folks in my network, I was able to get back off the ground quickly and get back to work. Every founder should have a roster of folks to call when things get hard. 

• Every solo founder needs a squad - To launch an app or anything really, you need to have a number of close people who are willing to test your app and a) give you the nitty gritty feedback you need to hear and 2) are willing to go above and beyond to debug issues with you. At Google, we had testing teams, or on my smaller teams, we had bug bashes where folks would hammer on a product before launch. As a solo founder, no one is financially incentivized to help me find issues in Wanderly. Thankfully, I have a few people in my network who were willing to test new builds at 10 pm at night, send me logs, write out a list of bugs with screenshots, and more. They’re heroes in my book, and I can’t imagine having launched this pivot without them. 🦸 🙏

So, from here, I have 2 new apps and a new go-to-market strategy, and I’m hoping things continue to grow from here. If nothing else, I’ve learned a lot and invested in the future. Now, it’d just be nice to have it pay off. 🤑

If you enjoyed this post, please help me out by downloading Wanderly from the App Store or Play Store and leaving a review or rating. 🙏

I’ve been working on a major go-to-market pivot for the last eight months, and it’s finally launched! This is part 1 of a two-part post about the pivot. This post will focus on the pivot itself. Part 2 will focus on my journey.

It would mean a lot to me if you reviewed Wanderly on the App Store / Play Store. Or, better yet buy a Wanderly gift card for a loved one! You can use the code RUNNINGTOWARDS for 20% off all book + subscription packages until the end of the month. Thanks so much for your support. 🙏

I just launched Wanderly’s biggest pivot to date: Wanderly is now a book publisher! Any Wanderly story can be turned into a hard-cover keepsake book with the child’s name and image right on the cover. 

I’m optimistic and excited about this direction. First, my daughters love their books. Alanna (now 6) got her first Wanderly book and was glued to it for 20 minutes straight when it arrived. Selena (now 2) has been too young to really interact with the phone version of Wanderly, so we made a book based on the types of choices we thought would fit her. It’s been her “favorite book” since it arrived, and she’ll eagerly tell us about her favorite pictures. 

Bringing Wanderly stories off the screen and into a child’s hands also lets me do some things that have been hard to do:

• Marketing is a LOT easier. It was always hard to market Wanderly; a lot of parents are anti-screen, it’s hard to get photos of kids using Wanderly that don’t make them look like screen zombies (even though their imagination and reading skills are engaged), and subscriptions can be hard to sell. With this publishing pivot, I can:

  • Market a physical book, which plays better with screen-cautious parents and feels more tangible to gift-givers

  • Expand my total addressable market (many friends who were previously lukewarm for screen reasons are now enthusiastic)

  • Create packages of books + subscriptions + add-ons, which give me more room for pricing experiments

• There’s now a tangible reminder to engage with Wanderly that’s physically in a child’s space

• I can do some pretty cool things with the art (more on that later)

There were a couple of key moments that convinced me that this pivot was the right direction to go. I’ve been getting consistent requests for printed copies of Wanderly stories since we launched, but I wasn’t sure how common that request was until I did a focus group with my daughter’s Kindergarten class. After spending about 30 minutes with Wanderly, at least 50% of their “I wish” statements were “I wish I could print out the story.” 

A week later, I met with Nick Baum, the CEO and founder of Storyworth1. He and I originally met at Google, but since then, he’s built a beautiful business that helps families document and print their stories. He pitched me hard during our meeting: I really needed to get into the printing books business. He had originally started Storyworth as a subscription-only service, and it was only once a friend pushed him that he printed his first book. Holding his own father’s stories in his hand, he knew publishing was the go-to-market strategy he'd been looking for. 

Nick’s pitch on the heels of my library focus group sent a strong signal: at the very least, I should try printing out a story. So I did … and Nick was right: having a book in my hands felt different! The images, which I had always gotten great feedback about, really popped. Putting Alanna's name on the cover and the spine made it feel like it was hers, not just words from an app. And I didn’t feel like my phone was the only way to experience Wanderly anymore.

In addition to being something you could hold, I can also do something really special with Wanderly images. For long-time readers, you know I’ve had to walk a tightrope between AI-image generation, safety, and the latency demands, opting to parent-review each Wanderly image and then do real-time matching. With a little extra offline time, I can transform a Wanderly story image into something really wonderful and personal:

An additional layer to this pivot is that Wanderly is now an app on the App Store and the Play Store! 🎉 When I started the app in late Nov ‘23, I thought porting Wanderly to the app stores would change Wanderly’s trajectory. Since it took far longer than the six weeks I was quoted (more on that in Part 2), I had lots of time to think about “Is this the answer?” I still think that porting Wanderly to a mobile app is the right call for my target audience and what can be enabled in the long term, but it didn’t solve marketing challenges. I think this publishing pivot feels correct for so many reasons. I hope I’m right. 😀🤞

Together with Wanderly on iOS and Android and publishing books, I hope the pivot resonates with potential customers. I’ll be reporting back in the coming months. 🫡

It would mean a lot to me if you reviewed Wanderly on the App Store / Play Store. Or, better yet buy a Wanderly gift card for a loved one! You can use the code RUNNINGTOWARDS for 20% off all book + subscription packages until the end of the month. Thanks so much for your support. 🙏

I’ve run three scrappy evals between OpenAI’s models and Anthropic’s models, which I wrote about in October ‘23 and May ‘24. Each time, I’ve been excited by Claude’s potential, but it’s fallen short of GPT’s overall capabilities. 

When I tested in May, I narrowed my evaluation criteria to a concept I’m calling “system-prompt2 drift”. To test system-prompt drift, you can ask the model to do something hard and measure if the model gets less good at doing the hard thing as the conversation gets deeper (and the model usually regresses to its default). In my case, that was trying to get the model to output a Flesch-Kincaid reading level between levels 2-3, equivalent to a 2nd-grade reader.

Below is the graph from my last newsletter. You can see that Claude 3 drifted away from the desired reading level as the conversation got deeper. Anecdotally, some of the stories with Claude 3 got off the rails. 

But that’s changed with Claude Sonnet 3.5! During my eval that was completed this weekend, you can see that Claude Sonnet 3.5 is in the green zone and stays in the green zone… even more consistently than OpenAI’s GPT-4o3! 

So I rolled out Claude Sonnet 3.5 yesterday to Wanderly. And for anyone who reads this who is at Anthropic, I wanted to say kudos! I’m sure a lot of love went into this model, and it’s paying off.

As an aside, it’s a real luxury to swap so easily between models quickly. In a few years, it’ll be interesting to see how the companies differentiate and how their market share intersects with open-source models. Maybe the capabilities of Claude 4 and GPT-5 will be enough for most devs… and when open-source models catch up, where does that leave the industry? 🤔

A couple of quick thoughts before I get back to bug-squashing and polishing the mobile apps:

• I love the texture of the Claude responses. Claude grasped my desired Wanderly style well (which I want to improve even more now that Claude respects it). It also added little flourishes like all-caps onomatopoeia like “BOOM” and “HONK,” which I didn’t ask for but work well for my target reader. 

• The evals are actually helping me debug my prompts now that my instructions are being followed more closely. For instance, you can see in the graphs above that my intro needs some tuning to get a little lower on reading level. 

• The dev tools for Anthropic / Claude are still pretty thin. Given OpenAI’s mindshare headstart, there are fewer blog posts about Claude, the docs from Anthropic could use more examples, and Amazon Bedrock has been slow to support Sonnet 3.5 globally. If you want to be a developer on the bleeding edge of tech adoption, you’re going to have to work harder to use Claude for now. 

• A reader4 of my last post asked me why I never really talk about Google’s Gemini models: I've done a couple of tests with Gemini, and maybe it'll get better, but it's never gone well. e.g., https://x.com/fearofpoets/status/1758908757516624098. I've never seen reviews that convinced me that it was worth putting in the coding effort to give the API a shot. When that changes, I'll definitely add it to the side-by-side matrix though :) 

Alright, that’s all for now. I felt like I owed y’all an update because I revisited this topic several times, and Anthropic flipped OpenAI for the first time. Big news! 

In closing, ping me if you’re interested in testing out either the iOS or Android app. I’d love your help getting it ready for public release. 🙂

So when Claude 3 came out in March, there was a lot of buzz, and I had to see if it was worth it. But instead of a manual analysis like October, I decided to put in the work: It was time to create some evaluation infrastructure. 

“Evals,” as they’re commonly called at Google, are structured analyses done at scale on any type of results. When I was working as the Product Lead for Google for Educations’s AI team, we regularly conducted evals to determine whether our conceptual understanding of homework questions and learning content was accurate. 

For Wanderly, I’ve decided to evaluate latency and the Flesch-Kincaid reading level for each model. Respecting a reading level is one of the hardest things for current LLMs to do since most of their training data was on content not intended for children. There are a lot of different instructions in Wanderly prompts, but reading level is my canary in the coal mine: If the LLM can respect reading level, it’ll probably respect my other instructions, too. 

Beyond product quality, respecting Wanderly’s system prompts is an important safety requirement. Left unbounded, LLMs will have characters blow things up, pray to God, and introduce family members that don't exist. While these story patterns might be mathematically average story elements, they can harm a child reading them if, say, the child isn't religious, has same-gender parents, or a parent has passed away. Reading level just happens to be my proxy for making sure an LLM respects these other instructions since I can assess every response’s reading level in a quantified way. These other elements happen less frequently and are less quantifiable.

So, back in March, I built out a series of scripts that run Wanderly prompts on the same set of inputs to the model and output some analysis. Unfortunately, this diversion cost me about a week or two, and the result was the same: GPT-4 was just a little better than Claude 3 for Wanderly. However, it set me up to run the same analyses as soon as OpenAI’s GPT-4o dropped, and I could confidently roll it out to production within 37 hours of GPT-4o’s announcement.

Learning from my Evals

There’s been an increasing amount of dialog about running evals on LLM products (for instance, “Your AI Product Needs Evals” is a very good read). What makes running evals so hard for LLM use cases is that the output can be random by design. In Wanderly’s case, I’ve cranked up the creativity meter (also called “temperature), which results in more interesting stories but more variability. 

Evals can be a combination of manual vs. fully automated, and I’m still very much on the manual side. But what was interesting about moving from fully manual → a little more automated caused me to solidify my "feelings" when testing a new model and ensure I wasn’t playing favorites. When I first tried out Claude 3 in March, I was immediately very optimistic (again, rooting for the underdog). The first page of each story just felt more aligned to the correct reading level… but then I realized that I was probably biased. 

Maybe by lucky accident, I also ran a story that forced me to get more rigorous: Instead of a simple "Help an Animal" story about helping a panda, a Claude 3 story diverged into Alanna convincing all the local factory workers to stage a protest until the factory owners decided to stop deforestation. While I would have been proud of Alanna for doing this, I did feel like the story was a) not at her reading or comprehension level and b) a spectacular divergence from the intended story premise. 

To counter my bias, I wrote scripts to take the same story seed and run it multiple times across each model. This type of analysis using the website would feel extremely tedious, and I would have stopped after a couple of gut checks. In addition to checking the first few pages, I also needed to evaluate the rest of the story and see if this political activism story was a one-off. 

This is where I started to observe system-prompt drift. Over the stories I tested with Claude, as the conversation got deeper, the system-level prompt with all the reading-level instructions got ignored in favor of (I assume) Wanderly’s most recent message to the model. This was a big red flag for me.

To put a fine point on it: If I can’t trust an LLM to respect reading level instructions, I can’t trust it to respect my other safety instructions. Anthropic has emphasized its unique constitutional AI approach as a safety measure; in fact, Anthropic was founded as a counter-approach to OpenAI’s safety approach (the first major safety exodus before yesterday’s announcement about Ilya et al.’s departure). Safety comes in many flavors, and I’m optimistic that Anthropic’s new product leadership will hold the line on overall model safety while also increasing developers’ confidence in output reliability (Hi, Mike! 👋).

So, after a couple of weeks of investment, I ultimately decided to keep GPT-4 as the base Wanderly model and shelf my analysis until I made a meaningful change.

Updating to GPT-4o

Fast-forward to Monday morning and OpenAI dropped its latest model: GPT-4o. The announcement was great news for developers like me: the GPT-4o API is immediately available, half the cost, and twice the speed. 

I swapped out a single line of code in my story engine and could immediately sense the speed difference. I also noticed an increase in coverage on my newly launched image-matching system.

Then I dusted off my eval scripts and, within several hours on Tuesday, verified that swapping to GPT-4o was safe to roll out to production during Wanderly off-hours Tuesday night (i.e. while children are sleeping and usage is low). 

Trying to get an immediate upgrade to Wanderly wasn’t top of mind when I did my eval investment in March. But almost instantly, Wanderly users get a higher quality experience, and I get a 50% discount on pricing. It’s nice when the adage “going slow to go fast” actually pays off. 💸 🚀

I’ve wanted to write a post about how I think about AI images + children’s stories for a while, and this felt like the right moment. So here goes!

AI image generation has come a LONG way, even in the last year. Here’s how Midjourney has progressed in ~2 years:

So, what does this rapid progression mean for Wanderly? My goal with building Wanderly has always been to create a child-led, open-ended space for children to learn and explore with their imagination. AI technology is a tool that enables this type of product in a way that’s never been possible. But that doesn’t mean I will try to shove AI into as many places as possible. 

I’ve wrestled a lot with how to incorporate AI images in Wanderly best: on one hand, it’s the first time we’ve ever been able to generate an on-the-fly image in 30 seconds or less using only words. On the other hand… I don’t completely trust AI image generators yet. At least not enough to go unsupervised with anyone’s kids. For this post, I will explain why I don’t trust AI image generators, what I think AI image generators are great at, and how I’m using (and not using them) for Wanderly. 

Why I don’t trust AI image generators with children (yet)

When I say I don’t trust AI image generators with children, I mean a couple of specific things:

• I have a few values that I think are important about images for kids. In essence, they boil down to “do no harm”:

  • Avoid images that reinforce1:

    • Racial & cultural biases

    • Negative gender stereotypes

    • Oversexualization

  • Avoid problematic misinterpretations of prompts

  • (And still relevant, although less so) Avoid body horror (e.g., “AI hands”)

• It is important that Wanderly be beautiful, but it's even more important to be safe. I believe this from a brand-risk perspective, but I also believe that knowingly exposing children to an unlikely but unacceptable outcome isn’t responsible. 

While these principles are an extension of my values, my experience with AI image generators made me leery of unsupervised AI image generation. 

When Midjourney v5 came out (March ‘23), I was just starting to add images to my Wanderly prototype. I vividly remember two experiences around image safety:

• Generating images of “Venus” (the planet) and 50%+ of the images included a nude woman

• Generating images of a friendly alien and finding myself with 3 images of an alien and 1 image of a dark-skinned girl. 😱

In both these cases, if these images had shown up in Wanderly (which I’m very glad they didn’t), it could have caused harm to the child reading the story and probably made the parent pretty angry. 

When I encountered these examples, I decided not to dynamically generate AI images with Wanderly stories (even though my competitors are). Instead, I launched Wanderly with images generated offline, matched to a page based on keywords. That means every image on Wanderly is parent-reviewed, which keeps the image quality high. In fact, several users have said Wanderly images are their favorite feature, and I’ve even received print requests. 

To Midjourney and DALLE’s credit, they’ve gotten a lot better (i.e., more “steerable”) in the last year. There’s also less body horror (i.e., most body parts are where you’d expect). When DALLE-3 came out last year, I even had a moment of existential crisis. Was I being too conservative? So I ran a few tests. It was much better, and the magic of dynamic images sorely tempted me, but there were a couple of things that kept me hesitating: 

• Multi-character prompts often result in chimeras (i.e., two characters blended into one hybrid character), and a key piece of Wanderly is the companions in your story

• Most image generators default to white characters (except for the hot minute when Gemini’s image models broke the internet). 

• Some key children’s characters are oversexualized, like fairies2 and mermaids. 

Beyond safety, I think a lot about product experience. I did a bunch of experiments where my story engine generated image prompts, which I passed to the DALLE-3 API. The images were interesting and often pretty, but covering all my bases with prompt engineering was hard. It was hard to control consistency, the AI generator often added things that weren’t asked for (like extra characters) and it was hard to control subjective things like “too scary”. With this approach, there was a risk of a young child getting confused and amped up before bedtime… and all parents know that’s not a recipe for product market fit. 

What AI Image generators are great at

But AI image generators do have their uses. Almost every image in Wanderly is AI-generated… but they’re also parent-reviewed (shout out to my husband / Wanderly Art Director, who does a lot of the image generation). We also spent a lot of time aiming for quality (Midjourney for most images, since it’s much better than DALLE-3 on consistency; i.e., once you find an aesthetic you like, it’s easier to replicate). 

In my opinion, AI image generators are great at cheaply creating beautiful images. If I wanted to build something like Wanderly without AI image generation, I’d need a massive budget for buying or commissioning art. There’d be no way I’d be able to build this business. I use AI-generated images for the in-app art, but I also use AI images for backgrounds for ads, stock photography, email headers, and the Wanderly logo3. I’m not opposed to using human-generated images either; in fact, I commissioned an artist for all of the Webster the Spider illustrations since I wanted something very specific.

I also like using AI image generators for inspiration. I used Midjourney to brainstorm Wanderly’s colors: I asked for color palettes for a kid's app, mined images for ones I liked, and then extrapolated my app colors from the image into my Figma mocks. I also like using AI image generators to mock up new features. In a typical Google design cycle, we’d do wireframes before getting illustrations. With AI image generators, ideating mocks with high-fidelity image assets is cheap and fast. 

AI Images & Wanderly

Which brings me back to Wanderly’s latest update. Before this week, Wanderly had about 200-ish AI-generated images of various settings or simple objects (e.g. jungle, space station, crown, ship, etc.) AI was also used to generate our avatar and pet icons:

Although Wanderly images have been a huge source of user delight, they have also been the biggest source of feedback. Keyword matching was primitive and often got things wrong (e.g., a “shark” instead of a “whale shark” or a “boat” that looked like a pirate ship when a child was reading a story about a “leaf boat”). I needed to figure out how to match images semantically to the story. 

With this update, Wanderly increased the total number of images to 300+ background images and 160+ characters and objects… all matched to stories using LLMs and natural language understanding. This new system also enables foreground images and allows characters to express themselves (generated using Midjourney’s new character reference capabilities). It also mirrors an approach used by many successful video games. 

An overlooked aspect of generative AI is that you really only get a certain “budget” of control over the output of these models. If you overspend your budget, you get artifacts, chimeras, or extra things you don’t ask for. It’s not a coincidence that most of the beautiful generative AI demos have simple prompts. 

The models just aren’t ready yet to do multiple characters, a contextually relevant background, and character consistency all at the same time, especially not in a way that’s safe for children. So Wanderly’s approach of parent-approved images + dynamic semantic matching allows me to get the best of multiple worlds: contextually relevant, visually consistent, and safe images across multiple contexts and characters. It comes at the cost of coverage, but that can be remedied with time. 

It’s not 100% where I want it yet: Coverage and matching aren’t perfect yet, but it’s substantially better than before. And I want to implement multi-character images… but I have a path towards making this all possible within the next few months. 😊

If you’d like to check out the latest update, head over to https://wander.ly. There are character expressions in the gray wolf, lion, and red panda in Help an Animal stories, with more coming soon. 

I also had the opportunity to chat with Tiffany Chin from Cocoy.AI about Wanderly, AI, education, and coding. Check it out here.

I wrote my first line of code for Wanderly about a year ago. As a career PM, I wanted to return to my engineering roots and gain the independence to create my own software projects. What started with a one-hour tutorial turned into what Wanderly is today. 

When I was leading Grasshopper, the team had a practice of occasionally screenshotting the app so we could look back at how far we’d come. It’s too easy to get lost in the weeds and forget to zoom out. So I started my own “time machine” on February 26th last year. 

I’ve been using my nascent social media video-editing skills to compile a walkthrough of Wanderly’s progress over the last year. Share it on social media if you like it (Instagram, Twitter or Threads)!

Big thanks to this community for helping me get this far! 🙏🥰 You continue to challenge me and offer me great advice, and I hope to repay it. I’m currently working on a major Wanderly image improvement, so expect an update on that in the coming weeks (and I’ll catch up on the info diet soon). 🏞️

Writing this up forced me to do a few things:

• Really understand my acquisition funnel. I knew this would be the first thing I would be asked, and I needed to be confident in my conclusions.

• Dig in on marketing. I’m a product person, and marketing is a growth area. This update forced me to get clear on target audiences and market size.  

• Reconcile my priorities with reality. Due to the chaos of the last few months, I had things I wanted to build but hadn’t validated my intuition with data. Now, I know a few things I need to prioritize differently. 

Although I initially wrote this up for a couple of close confidants, if you have advice or thoughts, I’d love to hear them. 

Summary

Wanderly is still very early, but there are signs of life. I’ll be working on retention and quality as my primary goals in the next 6 months. I’m optimistic that upcoming changes will cause Wanderly to hit a new inflection point. 🤞

Highlights

• 2 subscriber monetization highwater marks: Revenue and new paying subscribers per month (X).

• The mobile app is a week or two away from being code-complete. 

• Revenue doubled in Dec due to gift cards (~$Y in subscriptions and $Z in gift cards). 

Lowlights

• Retention is not great, but there are little to no retention techniques right now. I hope to address this with upcoming product improvements, mobile app notifications, and experiments with lifecycle emails. 

Requests

• I’d love suggestions for managing accounts and/or taxes as a very small business. 

• I’d love thoughts about which metrics to make my north star(s). 

• I’d love a brainstorming partner for refining my “What is Wanderly” pitch and messaging.

Favorite recent user quote

"My dyslexic, 4th grader struggles with reading and comprehension. She went from absolutely dreading reading to choosing reading over playing outside and then hiding under her covers after bedtime with books… all in about 3 weeks because of Wanderly!" - Bill

General Orientation

Since this is the first update, I’ll start with some context setting on vision and progress-to-date.

Wanderly: Magical Adventures for Curious Children

Wanderly is a story-driven educational app that encourages curiosity and imagination. Imagine choose-your-own-adventure stories, personalized picture books, play therapy, vocabulary-building stories, and mad libs all had a baby. This is made possible using Generative AI tools, specifically LLMs and text-to-image generators. 

My goal with building Wanderly is to build a self-sustaining business that inspires and educates children while also deepening connections with family. In using Wanderly with my own family, I’ve found a few things:

• It’s cooperative screen time that I don’t feel guilty about

• I’ve found out more about my child (her interests and her friends) than I ever have just asking her questions

• Wanderly provides a very cool springboard for imaginative play and conversation beyond the app

I hope that families that use Wanderly will have the following outcomes:

• Increased visibility into their child’s life and choices

• Increased connection with family members

• Increased educational outcomes as a result of vocabulary and information exposure

• Increased child confidence as they get to make choices and play in their own imagination landscape

How Wanderly is Different

Here’s a rundown of other AI-stories-for-kids products via the lens of their approach and one thing they do well. And if you have other products you think are worth looking at, please flag them. 

Overally, Wanderly is the only product that leans into kid-agency in storytelling and story quality. Wanderly really leads with the choose-your-own-adventure style storytelling; [redacted] is the only story system that even has an option, but it’s buried. Wanderly is the only product that adjusts the reading level for the child, and I can also tell from each story engine’s prose that little tuning has been done (there are certain tropes that OpenAI’s GPTs do by default that I’ve had to work hard to suppress, like overuse of adjectives and adverbs, trite summaries at the end of each story, etc.). 

It’s also worth noting that there’s a lot of churn re: the creators of these products. Older players have moved on or are creators who view these story creators as side-projects. Some new entrants seem to be taking it seriously with small teams. I think the reason is that, although AI-created stories seem initially interesting and you can get a demo up and running quickly, making *good* stories with AI (good enough that people want to keep coming back) takes a lot of elbow grease. As a result, there were a lot of entrants when it seemed easy; now that easy isn’t proving to be enough, those who were in it for a quick buck are exiting (that was definitely the case with [redacted]). 

Target Market & Sizing

Overall, I’m targeting a US market and kids aged 3-12 (with a light hypothesis that 3-8 might be the sweet spot, but I’m still learning). I have two narrower markets that I’m eyeing as starting niches:

• Homeschooling families

  • Why is Wanderly a fit? Many homeschooling parents have control over their choice of curriculum and tools. They tend to be oriented around tools that give more agency to their child while learning and are interested in culturally responsive education (top reasons for homeschooling), which is a sweet spot for Wanderly. Wanderly is child-led and open-ended, which are key Montessori principles, too. 

  • What is the market? Spurred by the pandemic, homeschooling is accelerating in popularity. There are about 4.3M homeschooled kids in the US, with ~2M in my target audience (source). Homeschooling families typically spend $500-2.5K per year on materials and extracurriculars (source).

• Learning differences families

  • Why is Wanderly a fit? Wanderly has gotten some very positive feedback from parents of children with dyslexia and ADHD. I mostly have anecdotal evidence here, but I hypothesize that 1) personalized stories and 2) short bursts of reading followed by interactivity are key to keeping their interest. Parents of these children are interested in finding any resource that will help their child practice reading skills, which can be challenging. 

  • What is the market? Research seems to indicate that 1 in 5 US students are dyslexic (source), and roughly 10% of students have been diagnosed with ADHD (source). 

Progress to Date

Wanderly started as a “let’s see how far we can take this” kind of project back in Feb 2023, and I kind of fell in love with it. Here’s a snapshot of what’s been built and how the product is doing:

• Technical assets - React web application (and soon-to-be iOS / Android app) using AWS Amplify to generate illustrated stories. Story text uses OpenAI GPT-4. Images are pre-generated using Midjourney and are matched to stories using keyword matching. Monthly subscription enabled via Stripe.

• Business assets:

  • XK registered users, but mostly dormant. 

  • ~Y monthly subscribers @ $15/mo, Z annual subscribers at $135/yr, and $X in gift card purchases during the holiday season. Y new subscribers this month (a new record), but several unsubscribes. So far, the investment is bootstrapped from personal finances, and I expect to continue that through EOY unless there’s a major growth inflection. 

  • I have deliberately decided not to put gasoline on yet (e.g. Product Hunt, reach out to Xoogler network, etc). I’m hoping to get retention figured out before doing that.

• Marketing assets:

  • Marketing site at https://wander.ly with a 25% user conversion rate for clicking through to the login page.

  • Facebook and Instagram followings of ~300 users, all paid acquisition.

  • I’ve seen the most interest from users of parents who have a learning difficulty (esp., ADHD and dyslexia) and homeschooling parents. There have been a few inquiries from teachers but little follow-through. 

  • I don’t think I’m nailing the marketing pitch yet. I think it’s easy to say Wanderly is a story app or even a story-driven educational app, but it doesn’t yet get to the heart of the child-led and open-ended aspects of the product. 

Metrics in 2024

I’ve been working to identify what metrics I should be tracking and driving forward. Right now, I know revenue is a long-term goal but not the key metric for decision-making. Retention and active users are a classic but might be too abstract / trailing indicators. 

Here are some things I’ve been looking at, and I’d love your thoughts on which of these metrics should be my north star:

• ~$X/user with Facebook ads to acquire a new user who creates their first story

• New user attrition in Jan:

• Story quality in January

  • Average story rating (only for completed stories): 4.62

  • Average story completion rate: 45%

    • First story completion rate: 32%

    • Subsequent story completion rate: 57%

  • Average stories/user: 1.88. Here’s how many they created in Jan:

    

What’s Next

For the purposes of a public newsletter, I’m not going to lay out my entire new feature roadmap. 🙂 What I will say is that I’m planning on launching a new story wizard and a native mobile app in the coming weeks, which should address some retention and onboarding issues (as well as some feature requests from my #1 user: my 5-year-old daughter). I’ll also be working on user lifecycle emails and cultivating a group of users to give me regular feedback. Beyond that, you’ll have to wait for the next few newsletters. 😉

While developing Wanderly, I’ve been mostly thinking about parents, but many people I’ve chatted with have encouraged me to think about people who might want to give Wanderly as a gift. Parents are saturated with child entertainment options and generally have a monthly budget, but relatives are always looking for novel gifts for birthdays and holidays. Wanderly already works well with aunts and uncles – for example, check out one of my favorite user reviews:

"So I’m on this 4 hr road trip from Little Rock to Dallas with a 4 yr old and a 7 yr old and Wanderly is the hero and blockbuster hit of this trip. I’ve now used it to tell 5 stories and they keep asking for more."  - Auntie Camille

The holidays are a perfect time to experiment with gifting: Almost everyone is in a gift-giving mindset, so targeting users and messaging are a bit easier to do. As Black Friday approached in November, I realized I would have to wait a whole year for these ideal experiment conditions to happen again… so I sprinted. I created a true MVP since I didn’t have time for “feature creep.” When I launched the gift card campaign on Cyber Monday, I hadn't yet finished the flow to redeem them! When the first person purchased, I was surprised, and then I rushed to my computer to polish things. 

In the spirit of building in the open, here’s what I built and what I learned. 

An MVP of Gift Cards

I developed a gift card concept around two products and price points: a $10 package for 10 stories and a $30 package for 30 stories. Each had a cute name: A Road Trip (since 10 stories could cover about 2-4 hours of content) and A Season of Stories (for 30 stories approximating three months at a more regular usage pace). 

As I started to build it out, I realized there was a major constraint: My whole payment flow is “in” the app. Without significant engineering investment, using my existing payment system would require every gift card purchaser to go through new user signup – too much friction1. 

Here’s what I did instead:

• Used Wix’s online store feature to have “physical”2 items to purchase.

• Leveraged my trial story infrastructure to add additional story credits to accounts when given a gift card.

• At first, I manually updated the story credits in the database, but later, I built a system with a separate gift card table that auto-redeemed story credits at login.

• Advertised Wanderly gift cards on Facebook and Instagram with gift-centric copy & targeted to grandparents, aunts, and uncles. I also sent a few email campaigns to existing users. These all funneled to a “Gift Cards” page on my Wix site. 

• Tried out coupons around Black Friday and the last week before Christmas. 

• Made a few printable gift cards to add some legitimacy and delight to the gift-giving flow.

• Used Google Apps Script to send personalized emails to users to notify them of their gift card.

What I learned about gifting and Wanderly

I had a few challenges during this 1st gifting season: 

• The product is still really early.

• The marketing messages are nascent; I don’t have a lot of testimonials or video content.

• The gift card landing page is unoptimized.

• Oh, and life was still crazy with a 5 and 1-year-old at home and unwinding my dad’s estate during the holiday season.

But despite all the challenges, I earned as much revenue from gift cards in December as I did from the regular product with subscriptions. So, I think there’s some potential here for Wanderly as a seasonal gifting product. 

Here’s some interesting tidbits from the experiment:

• Some of the gift cards were for people who had created accounts months ago (a great sign of stickiness).

• Most purchases were in the week after Cyber Monday and the week before Christmas. 

• $10 and $30 gift cards were about equally popular.

• My volume was too low to know if coupons made a difference.

• I also implemented it completely wrong: I leveraged my existing infrastructure too much and defaulted to redeeming on an email address. I should have used coupon codes. Next year.

• Only 50% of my gift cards have been redeemed as of writing, but I’ll send a nudge in a couple of weeks.  

It was also interesting to see people's willingness to buy story credits via gift cards. I’ve avoided story credits as a monetization model so far because I wasn’t sure what a “story” would cost me. But now that I’ve found that the sweet spot for Wanderly stories is 5-10 pages3, costing in the per-story unit doesn’t feel constraining. Sometime in the next few months, I’ll likely create a middle pricing tier that’s a bit more pay-as-you-go but doesn’t include some of the more expensive Premium features. 

In the meantime, I’ve unlaunched gift cards. Because my MVP was implemented in a way that requires a lot of manual work (i.e., “things that don’t scale”), I’m being protective of my time and focusing on other experiments. I’ll definitely bring gift cards back next holiday season or earlier, but with a new flow that builds on my learning here. The nice thing about building my MVP with Wix is that I don’t really need to worry about maintaining much of it while I continue to iterate on core Wanderly.

I’m almost done working with a Toptal engineer on an iOS & Android version of Wanderly, which is also causing me to revisit my marketing site and branding (like app icon and splash pages). When that arrives, I’ll do a long-form post about the process, what tools I used, and any early learnings. I’m approaching one year since my first line of code for Wanderly, and I’ll be doing a fun reflection on that, too. 🙂

Hoping you all have had a happy and healthy start to 2024, and I hope this year is better than the last!

If you’d like to support things I care about, consider donating to the Oakland Literacy Coalition to improve literacy in my hometown, or buy a Wanderly gift card for a loved one (use RUNNING for 25% off). Happy Holidays! 🎁

I think 2023 was the hardest year of my life: I lost my father, my five-year-old didn’t have reliable childcare for about 4 months, my one-year-old is… one, and I founded Wanderly, a story-led educational app using generative AI. But I’m probably the happiest I’ve ever been about how I’ve spent my time and energy. 

As the year closes, I want to return to where this year started: In January 2023, I declared an intention to align my goals with my values. In the same post, I announced a new Etsy shop1 and decided to move on from explorations in crypto. Progress isn’t linear. 

But I also set up some hypotheses that I wanted to test and set an intention to come up with guiding values. The mantra of “aligning my goals with my values” rang in my ears all year, and it steered my path even when things got too crazy for me to be deliberate in my actions. So here’s how I did this year. And many thanks to all of you who have been with me (and guided me) on my journey this year. 🙏

3 Hypotheses

Here’s how I did on each of the three hypotheses I set out to test this year:

• 🟡 I can be content not pursuing vanity metrics - I’m pretty content with Wanderly’s progress, but that’s because it’s nascent, and I’m not pursuing growth yet. But I often look at numbers of users, sales, or funding inquiries for validation beyond what’s useful to drive Wanderly forward. Conclusion: Medium successful.

• 🟡 I can make money on my own - I’ve made some money from Wanderly this year. By no means is this enough to pay the bills, but I also don’t think any new company comes out the gates with a firehose revenue stream. A wise mentor told me that revenue can be a longer-term goal, but it’s too early to set it as a key metric; if I elevated revenue to a key metric, it would likely skew Wanderly into something I don’t like. Conclusion: Yes, but not at scale… yet. 

• 🟢 I can build a technology solution on my own - Much to my surprise, I’ve built and grown Wanderly using my coding skills (with some help from great tools and ChatGPT). I’m very excited about the foundation I’ve built here and about continuing to grow my technical knowledge and skills. Conclusion: Yes!

4 Guiding Values 

In January, I said I would establish guiding values, publish them, and evaluate my alignment with those values. I never did… until now. Better late than never. 🙂 

💗 To Love

So much of this year was about family. I’m proud of how I supported my dad in his final year of life. I’m proud of how my husband and I worked together to support our children through tough moments of growth. I’m proud that I still found time to invest in my marriage and supported my husband through a career change. Even though it cost me some professional momentum, I don't think I'll ever look back and regret supporting my family this year.

🛠️ To Build

I worked with a fabulous coach who helped me identify that I’m a builder. And I decided to build Wanderly. It started as an experiment to return to coding and gain independence. It quickly turned into an all-encompassing project where I practiced engineering, marketing, UX design, and business management. I feel more confident in my identity as a builder than ever, and I want to continue to lean into it. 

📚 To Learn

I’m a happier person when I’m learning… and wow, I learned a lot this year. I am learning how to code (again), how to advertise on social media, how to monetize, how to make a TikTok, and how to tell great stories. And I still have so much more to learn, and I also made many mistakes that have taught me lessons. I’ve also learned a lot from the generosity of this newsletter community. 🙏 

To learn also has a double meaning: Learning is the core of what I’m trying to build with Wanderly. Stories have so much to teach us and can be a delightful surface for passive learning. 

💪 To Empower

As a college RA, a manager, a parent, a board member of a non-profit, and a founder, there’s always been a throughline to my work: I want people to be the best versions of themselves. 

This year, my children learned skills to express their needs (and I learned skills to help), I joined the board of the Oakland Literacy Coalition, and Wanderly told stories where each child was the hero of their own story, and they got to decide what happens next. I hope to turn Wanderly into a tool that superpowers each child’s sense of self and imagination.

Overall, I think I spend very little time outside of these 4 guiding values, and I look forward to carrying them with me into 2024. 

5 Mistakes

In the spirit of humility and learning, here are my top 5 mistakes from 2023:

• Not doing due diligence on naming - I should have thought harder before launching my code name into a public-facing name. 

• Being glib about content strategy - At the beginning of Wanderly’s journey, I naively thought that I didn’t need to have a content strategy because there were infinite stories due to AI generation. But people still need to be hooked by content before they are willing to spend their time. 

• Not understanding ad fatigue - I got lucky that my first few social ads worked well, so I was surprised when they became less effective over time. Then, I learned about ad fatigue. Now, I’m experimenting and learning more with better results. 

• Not cultivating a core group of users for feedback - There was a lot going on this year, but if I had to do it over again, I’d invest more effort in cultivating a core group of users for regular feedback. I’ll be kicking that into high gear early in 2023. 

• Stressing too much about other people in the space - The space of creating kids’ stories with AI had many players and no real winners this year… but I stressed out each time I heard about one. I wish I’d spared myself, as I don’t think stressing helped. And in some cases, I didn’t need to worry at all. For instance, the app Stori announced at Google I/O shut down last week (only 2 months after public availability).   

6 Hopes

I’m hopeful that 2024 will be better for me, and I’m also hoping it’ll be better for the world. There’s too much going on right now that makes me sad. I try to focus on and amplify the good (check out Future Crunch or the Progress Network if you haven’t) and hope it counteracts the ugliness. 

Here are 6 hopes I have for 2024. I hope…

• My family is happy and healthy, and I wish the same to you and yours. 

• Wanderly continues to grow. 

• To keep learning curious things and to talk with passionate people. 

• To continue to grow and invest in community, including this newsletter.

• That AI opens more doors than it closes.

• For peaceful resolutions to the many tensions in our world. 

Thanks for a great year, everyone. I appreciate your support and accountability as I venture off into uncomfortable new things. I hope your new year starts well, and I’ll be back in January with many fun updates. 🎉

🎁 A Wanderly Gift Card promo: Use RUNNING for 25% off all story gift cards.

In this post:

• 🎢 The last month of OpenAI news… in emojis

•  🎉 Wanderly Wins

• ℹ️ An action packed info diet

First, thank you so much to everyone for your support on my last post. It was personal and vulnerable, and I was touched by all the support as well as others sharing their own stories. And if I haven’t responded, apologies. There’s been a lot going on. 🙏🙏🙏

I’ve wanted to write a newsletter for the last several weeks, but things didn’t feel steady enough to match my writing speed. For those of you who aren’t following the tech news closely: OpenAI’s Dev Day was a blockbuster event announcing a TON of new changes, then less than two weeks later, Sam Altman, the CEO of OpenAI, was ousted by OpenAI’s non-profit board. In probably the strangest week of corporate Game of Thrones, Sam was back within a week.

As an entrepreneur leveraging OpenAI’s tech stack, the last few weeks have been… strange. Here’s roughly how it felt, in emojis:

• 🥳 Nov 6 - Dev Day - Wow! So many toys! And the API I use will be faster and cheaper!

• 🤯 Nov 8 - Whoa! This Assistants API is amazing! I should use it to rewrite my whole backend. 

• 😰 Nov 9 - Oh shit. This Custom GPTs thing is gonna eat Wanderly’s lunch. 

• 😮‍💨 Nov 10 - Never mind, Custom GPTs can’t support what I’m doing yet. 

• 🤔 Nov 11 - Am I missing an opportunity to leverage Custom GPTs as a marketing channel? 

• 👩‍💻 - Nov 12 - 16 - Maybe all of this distraction for now; I’ll wait for things to mature a bit since Assistant API and Custom GPTs are missing features I need. Keep building.

• 🤯 - Nov 17 - Holy shit. The OpenAI board just fired Sam Altman!

• 🍿😰 - Nov 18-21 - What is going on?? Why did they fire Sam?1 If all of OpenAI resigns, does that set back the industry for at least 6-12 months as everything gets rebuilt?

• 🫡 Late Nov 21 - Sam’s back. Keep calm and carry on. 

For the curious, here’s my take on the big drops from Open AI’s Dev Day:

• Assistants API: A still nascent API with lots of potential. The idea is that it will handle a lot of conversational threading on behalf of developers + automatic retrieval (you can upload a lot of context, and it will handle embeddings for you). It's very cool / a big engineering timesaver. However, the retrieval tech is still early, and the API doesn’t support streaming yet. So I’m gonna give it some time to bake. 

• Custom GPTs: I’m still pretty mixed and unsure on this one. Any ChatGPT Pro user can create their own custom agent with their own rules. Eventually, there will be a CustomGPT marketplace, which is now launching in January. There’s a lot to unpack here: 

  • It’s a new feature, marketplace, and social advertising lever all at the same time. It’s a great way to get ChatGPT Pro users to 1) use ChatGPT more, 2) become creators and then advertise to their network, and 3) create another value add to being a ChatGPT Pro user. 

  • Someone on Twitter had an amusing observation that maybe it’s just a ploy to get around copyright / get more training data with all the PDFs people will upload to power their CustomGPTs. 🤔 

  • With an 8K character prompt limit, it restricts what’s possible. I tried replicating Wanderly and couldn’t get it to the quality bar that I think reflects Wanderly. But it is forcing my hand on how to handle Wanderly images. I’m unsure if I should make a CustomGPT to advertise Wanderly. Time will tell. Let me know if you have any opinions! I’d love to hear them. 

All this is to say… it’s been a long month since my last update. And on top of all of this, I’ve been dealing with the aftermath of my father’s passing. Since the memorial, it’s been its own rollercoaster: taking art off the walls, finding evidence of little and big ways he loved me, a ton of paperwork, finding some less fun things, and then just having to laugh about it all. 

So, I’m going to take a moment to celebrate some of the things I’ve been able to learn and make progress on with Wanderly. Gotta count the wins. 🙂

• It was fun that my AI will not kill Hollywood post got shared on Benedict Evan’s newsletter. 👋 to all the new subscribers!

• I rewrote my Wanderly storytelling backend to leverage that GPT-4 is faster and to prepare for some new updates. I think the stories are even better. 

• I started experimenting with DALLE-3 API for Wanderly images; it’s more steerable than any AI image generator and just might be safe enough to deploy live. I’ll probably do a long-form post on that in early 2024. 

• I improved my onboarding flow after discovering I was setting the wrong default reading level during setup.

• I also hired a React Native dev via Toptal to start porting the app to iOS and Android. I think that’ll unlock some new opportunities and is a top feature request. I’m hoping for a v1 by the end of the year. 

• I launched Wanderly Gift Cards (and you should buy a gift card for loved ones this year! It’s $10 and $30 for 10 and 30 stories, respectively, and with the RUNNING coupon code, you’ll get 25% off). Although my revenue is small, in November, it was 50% more than in October due to gift card revenue. 

• I got cold-contacted by a major VC firm, which is fun. Note: I’m not currently raising funding, but I’m contemplating it for 2024.

Thanks again for all your support. I’m looking forward to getting back into building and writing in the coming months and hoping the progress on Wanderly accelerates. 🙂

Info Diet

I was mostly glued to Twitter for the weekend of OpenAI drama, but in addition to that news, here are some fun things that crossed my path:

• Industry news:

  • Biden announced an executive order on the development of AI models and products. (A quick summary and a good Hard Fork discussion). Surprisingly bullish on AI in Education, but introducing regulation for large model creators. 

  • Anthropic launches Claude 2.1 - 200K token context window, improvements on hallucination, and pricing changes. I saw this excellent analysis on the performance of the 200K context window (spoiler: not all context tokens are created equally). But then Anthropic saw the analysis and published some prompting tips.

  • Sundar/Google tweeted a video of Gemini, a cool video recognition + chat experience that was impressive but failed to demonstrate a practical application. 

  • A cool demo reel from Pika’s latest AI video launch. 

• General AI stuff (notice how most of these are about ChatGPT? It’s just the center of gravity right now). 

  • ~10% decrease in freelancer earnings 5 months after ChatGPT launch

  • A very thoughtful explanation of how the big AI models “know” things

  • Some CustomGPTs that have gone viral

  • Some researchers found a way to hack ChatGPT for training data, and now it’s a terms of service violation. 

  • A 90% successful strategy for jailbreaking ChatGPT

• Some AI-themed fun 🎊

  • Some deep fake fun with Taylor Swift and Ron DeSantis

  • A single creator made a 3-minute movie with AI, and it’s much better than the majority of the AI movies I’ve seen. It does a really interesting job of storytelling and working within the constraints of the current tech. 

  • “Make it More” has been a recent AI meme/trend. A friend launched makeitmoremoremore.com if you want to play.

But if you’re here because you are interested in my authentic journey, the “real me,” I’m sharing what I’ve been working through the last few weeks. 

My dad, Neil Holmes, passed away on October 18, 2023. Some things that bubble up for me:

• I haven’t fully processed my grief, but writing helps. 

• It was a good death. Over the last year, my dad and I were able to move beyond our complicated history and celebrate the things we have given each other. He passed as I held his hand. His caregiver and I were reading him Pablo Neruda, one of his favorite poets, as he took his last breaths. 

• I’m in awe of every end-of-life caregiver out there. To love someone and then to deal with the reality of losing them (both before and after the actual moment) is an incredible amount of emotional, intellectual, and often physical work. 

• My father loved beautiful things, and he shared that love with the people around him. His perspective on beauty has deeply influenced my life and my work. Allow me an opportunity to pass along some of it here. 

Neil found beauty in physics, where he spent many years researching the matter of the universe. During my childhood, we’d go for long drives, and he’d tell me stories about the stars and the sky. The scientific world knew him for over 250 publications1 and for founding a facility in Nevada that researched the nature of the U.S. nuclear stockpile without needing full-scale tests. He was featured in a BBC documentary for his work on iron at the Earth’s core, so just imagine this guy (@ 17:50) teaching you about science growing up. 🙂

I remember him most for sharing with me his love of food, nature, photography, music, movies, and words. Since I was 5, I saw my dad every other weekend. After time apart, he’d often greet me with a new, beautiful thing he wanted to share. I remember when he returned from a business trip with a copy of Haroun and the Sea of Stories in his hands, bursting with excitement. “I read the first page, and then I couldn’t stop; it was so beautiful. You’re going to love it.” And then he read me the whole book, including many funny voices, every night I visited him for the next few months. It was the first book I read to my now husband. 

To honor that part of him, I wanted to share some things that he loved and introduced me to. If some of these things are new, give them a shot; they’re great. If they’re old favorites, maybe this is a reminder to revisit. I know he’d love it if you felt just a little more awe or wonder in your life. 

• Nature - Yosemite, Piute Pass & Lake Sabrina, Mono Lake, the White Rim Road, the Alabama Hills, Moab and wildflowers, birdsong, rocks, trees, and camping where the Milky Way is the only nightlight you need. 

• Music - Going Home by the LA Four, St. James Infirmary by Louis Armstrong, Scheherazade, A Meeting by the River, Endless by Keith Jerret, M. Marais variations "La Folia", Rusalka Op. 114 Song to the Moon

• Movies - The Princess Bride, Indiana Jones, Star Trek (TNG)2, Star Wars (Original Trilogy), A Room with a View, The Fifth Element, The Lord of the Rings, Harvey, Little Shop of Horrors, The Sting, Strictly Ballroom

• Poetry - Rumi, TS Eliot, Pablo Neruda, Shakespeare, Break the Mirror

• Fiction - Haroun and the Sea of Stories, Moomintroll Midwinter, The Hobbit, Sabriel, The Left Hand of Darkness, Enders Game, Book: A Novel, The Monkey Wrench Gang, Desert Solitaire, Letters to a Young Poet, How to Live: Or A Life of Montaigne in One Question and Twenty Attempts at an Answer 

Losing a parent has also made me reflect on my mortality, decisions, and the uncertainty of life.  My dad loved his work but also dreamed of retirement. He retired after 42 years at Lawrence Livermore National Laboratory in November 2019… 4 months before the pandemic started. As an asthmatic, he took quarantine seriously; we didn’t hug, he didn’t travel to the mountains as he’d hoped, and he saw few friends. I see the gap in our photo history together, but also a happy reunion with my eldest daughter after vaccination. Where I am in my grief, it’s hard not to regret prior decisions. What I can do is focus on the end; it was good. I can’t regret a good ending. 

In the coming months and years, I’ll be grappling with my grief and how to honor him best. For now, I think it’s about making decisions that maximize joy and time spent with my family. I’m grateful for my job that’s allowed me to be so present these last few months. And thanks to you all for reading this small tribute to him. I miss him, and I appreciate you helping me fan the flame of his existence just a little longer.

A few months ago, I came across a Reddit thread called “Claude is criminally underrated” and a tweet talking about how Claude can better replicate an authentic-sounding “voice”. And so I started exploring Antrhopic’s suite of models…

Wanderly needs to tell educational, immersive stories

The goal of Wanderly is to tell stories that are so good that a child learns while they are entertained. With so many demos of large language models telling children’s stories, you might think that achieving this goal is easy, but it hasn’t been. 

The hardest problem to solve has been getting an LLM to respect reading levels. Children (especially those learning to read independently) need to read at a challenging, but not discouraging level. Because LLMs are mostly trained on documents written for adult audiences, they naturally tend to sound like they are writing for an adult audience. I’ve had to do a lot of prompt engineering to make sure Wanderly’s stories use the right language for children. 

For instance, GPT 3.5 loves adverbs… Even when I ask GPT 3.5 for simple text for a Kindergartener, it’ll often throw in a “suddenly”, “mysteriously” or “mischievously”. Then, the text complexity would drift higher as the stories got longer. This bothered me, especially as an educational product. I want more consistency for my users. 

GPT 3.5 only followed my story structure guidance ~85% of the time. When it didn’t follow my instructions, the plot of each story got a little … weird. There are “temperature” or “chaos” parameters that can make responses more predictable but also make them less creative. 

Lastly, occasionally, I’d get weird artifacts in GPT 3.5’s responses that would break the reader’s immersion in the story. For instance, at the top of a page, I’d get something like “Alanna just decided: ______” instead of integrating the choice seamlessly into the story.

A subjective framework for objective assessment

When I started playing with Claude, it initially just felt better. In retrospect, this was a red flag for my biases creeping in. But, after a couple of encouraging test queries, I decided to start the engineering work to fully integrate Anthropic (partially because there’s no Claude API playground like OpenAI). 

After getting everything working, the cracks started to show. When you view an LLM’s output at scale, you realize you can’t cherry-pick your favorite responses (like you see all over Twitter). Every response is a dice roll; some are good and some are bad. You have to evaluate responses in aggregate, ideally with a large sample set. I needed to find objective criteria1 to decide across all available LLM providers.

So, I came up with a framework. Based on Wanderly’s needs, I decided to evaluate models based on language level reliability, latency, story plot adherence, and story immersion. 

Here’s how each one performed:

The first thing that might jump out at you is: “Wow, Claude 2.0 was crushing it on a lot of things, but why is there a weird N/A in that last column?” The answer is Claude 2.0 was doing some amazing work, but this model has a fatal flaw for Wanderly: It reallllly wants to be an assistant, not an API. That means that every response started with “Here is your ______” or “Sure, I’d be happy to ____.” I was terribly frustrated by this. I tried giving Claude 2.0 instructions to remove the preamble, but it changed the prompt so much that it degraded reading level quality2.

Claude 1.2 was actually a huge improvement over GPT-3.5, and so was GPT 4. Claude 1.2 is also cheaper and faster than GPT 4. But there were two things that really put the nail in the coffin for Claude 1.2: 

• It sometimes added artifacts like adding the title of the story on the first page.

• It hallucinated some science facts3. 

I also found GPT-4 doing some nice, creative things in the story prose while still adhering to my story structure guidance. That was a cherry on top of my final decision to go with GPT-4 for Wanderly’s story engine. 

Anthropic’s models are impressive, but Claude is a weak product

So, while I think Anthropic’s models are very close in quality to OpenAI’s for children’s stories, this post would be incomplete without covering the product and business providing the models. Claude’s API felt incomplete and underresourced:

• The developer and support ecosystem is thin and difficult to navigate: 

  • There aren’t a lot of code examples of Anthropic API online. By contrast, OpenAI has a neverending forest of YouTube and blog tutorials. 

  • The Anthropic Discord is very quiet. I encountered one Anthropic employee who was extremely helpful but didn’t encounter a community helping each other.

  • Many infrastructure tools don’t fully support Anthropic (like Helicone, my favorite infrastructure tool; side note: it’s amazing)

• Anthropic’s console lacks an API playground and clear communication about payment.

• It took 3 emails, 2 weeks, and a ping on LinkedIn before I was able to upgrade my Anthropic API keys to a non-test environment.

Overall, I was surprised by how well Anthropic’s models did given how less popular it is. The biggest flaws seem like product flaws more than model flaws. I sincerely hope that the team at Anthropic will listen to developer feedback and build more flexibility and tools in the coming months. Then, I fully intend to return to this choice and would be delighted if the underdog won. I look forward to Anthropic being a formidable competitor, pushing forward the industry. 

A side note on scalable evaluation

This whole analysis took way too much time. I’m glad I did it and plan to do it again, but it brings up a hot topic in the space of Generative AI: evaluations. How do you know if a model is better? How do you know if a prompt is better? How do you know if your fine-tuning is working?4

Right now, it seems like a lot of builders are doing exactly what I’m doing: manual inspection. Some tooling systems, like Helicone and Freeplay.ai, are trying to build test suites, user feedback reporting, and A/B testing frameworks. I’ve even heard of folks using LLMs to evaluate LLM responses, which I’m probably going to try next time. But the answer is: No one has figured it out yet. What a time to be building. 😅

A Wanderly Update

Other than rolling out this major story engine update, I’ve made a couple of small changes:

• By popular request, added an Annual plan. I found this awesome writeup on how to think about annual pricing, but since I don’t know my average customer churn yet, I chose a 25% discount and we’ll see what happens. 🤷‍♀️

• I put multi-child profiles behind a paywall since it really only makes sense if you’re going beyond 5 free trial credits and it was my #1 question re: Premium plans. 

• I put an interactive demo using Storylane on the Wanderly marketing page, and I’m tracking to see what impact it might have. 

I also just got access to DALLE-3 less than 24 hours ago, and it’s more amazing than I was hoping. I think it’s going to unlock some magical things for Wanderly. Stay tuned for fun!

Info Diet

• There’s been some discourse about whether LLM wrapper startups (like Wanderly) will fail or thrive. Two different takes: The Information and a thoughtful tweet. 

• Anthropic is releasing research about understanding how LLMs “think”

• Base models make a difference in the ability to self-improve on a task

• Do LLMs have a sense of space? 

• Image input for ChatGPT opens up new avenues for prompt injection attacks

• If you want to enjoy something completely tangential, here’s an amazing visualization about how Daft Punk sampled another song to create One More Time.

To be honest, I've gotten so much farther than I ever thought possible. But I wouldn't have been able to get where I am without ChatGPT. So today I thought it might be interesting to share how I’m using ChatGPT to enable solopreneurship.

Note: I only ever use ChatGPT with the GPT-4 model, which requires a ChatGPT Plus subscription of $20/mo. For me, it's totally worth it.

There are two major things ChatGPT helps me with: coding and marketing. With coding, it’s magical: I ask ChatGPT to write snippets of code, write scripts, find bugs, and navigate productionization. With marketing, ChatGPT can generate a lot of ideas, but I have to rely on my own judgment to make something good happen. 

ChatGPT can help with marketing but has limitations

I’ve used ChatGPT as part of my workflow for all parts of Wanderly marketing: website copy, ad copy, creating a tagline, and rebranding Wanderly. A couple of things to note:

• It has access to all culture, which is helpful when asking for inspiration

• It has its own copy style; by default, it uses lots of adverbs and a large vocabulary, and most naming exercises usually sound like bad startup names.

• By default, it usually assumes that you’re a large company with lots of resources.

• It never runs out of new ideas and is always willing to engage with all of your ideas.

A typical brainstorming session usually goes like this (sample prompts in footnotes1):

1. I describe my vision for Wanderly, in as much detail as I can. 

2. I frame the marketing task I’d like to do.

3. Say “Before you start brainstorming, ask me some questions so I can provide more context”. 

4. Answer the questions, and then tell ChatGPT to start brainstorming. 

The thing about using ChatGPT to help with creative tasks is that it always defaults to the average advice… but marketing isn’t about finding things that are average. Marketing is about finding a unique spark and cultivating it. But the average advice is a lot of good advice. So I’ve used ChatGPT to remind me of table stakes and to get inspiration, e.g., think of SEO, referrals, and synonyms. Ultimately, I have to be the one shining the light on valuable things. My final marketing deliverable is usually something inspired by my conversation with ChatGPT (never verbatim), but I got there much faster due to its help.

ChatGPT is like having an experienced, patient pair programmer on my engineering team

For the readers who aren’t familiar with software engineering, there’s a practice called “pair programming” where two engineers work together on the same code at the same time. It’s often used to help engineers come up to speed or when working on a particularly gnarly problem. 

With ChatGPT, I have an experienced pair programmer constantly at my side: it writes code instantly, knows many industry tools and libraries, proactively suggests improvements, and it’s good… most of the time. Here’s my basic workflow (sample prompts in the footnotes2):

1. I describe my tech stack, to give the AI relevant context, e.g., I’m writing a React app.

2. I give it a large chunk of my code (I’m talking up to hundreds of lines of code and/or an entire stack trace)

3. I describe the change I want to make or the problem I want to solve

4. I ask for it to return example code

5. I copy-paste the code, run it, and then ask for improvements

6. Always say thank you… just in case ;)

It’s pretty wild. One of the first things I noticed is that coding with ChatGPT almost completely erased my imposter syndrome. One of the reasons I became a PM instead of a software engineer in my early career is that when I found a bug, I often felt like it was *me* (instead of thinking that it was the difficulty of the problem I was trying to solve). Now I can power through solving problems; my only barrier is willpower and patience. 

It’s not all 🌈s and 🦄s yet. I have an incredibly tough time using ChatGPT to help me with development environments and production deployment challenges. It’s still far better than trying to figure it out on my own, but ChatGPT can send me on wild goose chases. When I find things getting too weird or difficult, I start from scratch and initiate a new conversation with ChatGPT; if that fails, I’ll go to help forums like GitHub and StackOverflow. I’m not entirely sure what makes this type of problem so much more difficult than my other coding-related requests, but I hear many professional engineers bemoan this aspect of software development, too. Some things still require a decent amount of human-level intelligence. 

As I’m writing this, I have two feelings: 1) I know how incredibly quickly this post will age, and 2) I can imagine the future of coding with AI. My prediction: in 5 years, I’ll be able to build an entire product mostly conversing with AI. I’ll be able to feed my entire codebase into an LLM as context, and then ask it questions and change code. But also: I’ll be able to come up with a product idea, develop a business plan, brainstorm wireframes and high-fidelity user interfaces, and develop code… all by just talking to an LLM the same way I’d talk to a team at work. 

In order to make the best products that surprise and delight, I still think there will need to be a human with a vision piloting the output of these AI tools, but maybe that’s just wishful thinking. There have been whispered rumors of OpenAI having achieved Artificial General Intelligence (AGI), and maybe an AGI in five years will be able to build products far better than I can. But until then, I’m going to keep trying to build and manifest my vision for what I believe should exist in the world, and hopefully, I can use these tools to accelerate it.

I still miss my team 🥹

So, does all of this talk of ChatGPT and solopreneurship make me feel like I don’t need a team: Absolutely not. I really miss having a team. Thankfully, my husband is a great sounding board and support. But it’s not quite the same. A team can shine light in blindspots and provide the experience and taste AI can’t offer yet. And nothing beats the camaraderie of a group of people all working 100% towards the same thing. 

I hope I’ll be in a position to hire a team soon, but until then, ChatGPT is allowing me to de-risk Wanderly while keeping my burn rate extra low. 

So how is Wanderly doing?

About 6 weeks ago, I rebranded Wanderly. It was hard and I was really nervous, but it’s gone so much better than expected. My CAC dropped 25% on social ads, I’ve gotten a lot more questions about the product, and the comments are much nicer. Wanderly just crossed over 1K stories told, with an average star rating of 4.5. I launched a new story type: Help an Animal. 

I’m learning quickly, but there’s still lots of work to do. I’m experimenting with coupons, some social media promotions, and marketing site improvements, and I will also be working on improving the first-story experience. 

Share Wanderly with a friend

Info Diet

• Some hopium for the soloprenuer: Being patient and disciplined can pay off.

• OpenAI is on 🔥:

  • They announced DALLE-3 in ChatGPT (demo video + Midjourney v6 vs. DALLE-3 comparison)

  • ChatGPT can now see, hear, and speak

  • TBH, these announcements are 50% intimidating and 50% exciting because I’ll be able to use these APIs. I keep telling myself that it’s one thing to do a demo video of a one-off children’s story, but it’s another thing to productionize and refine AI-powered children’s stories that people want to return to regularly, and that these demos are functionally an advertisement for Wanderly’s use case… but I only convince myself half the time. 🙂

• ~20% of Americans use ChatGPT, and for many, it’s not part of their core workflow

• Amazon to invest up to $4B in Anthropic, tying these two companies together similarly to how Microsoft and OpenAI are tied. The open question on everyone’s mind is: What happened to Google’s $300M investment in Anthropic?

• Another good Hard Fork podcast about Bard Extensions and facial recognition.

• Feeling like humans weren’t made for the endless grind of capitalism? Watch this 30-minute video on the origins of the 9-5, how medieval workers used to work, and how the start of the workday was shaped by a bargain between a businessman and a church. 👀