Tina He - Building for Joy: Playful AI Development

Key Insights

“Tokens don’t lie” - The usage patterns of AI models and developer tools reveal what actually works better than any marketing claims. Context7 and OpenRouter token allocation data shows real-world adoption.
Play is the point - While 90% of AI usage might be productive work, the remaining 10% for creating joy and fun experiences represents a fundamentally new form of self-expression and socializing, now cheaper than buying someone coffee.
Writers make natural vibe coders - There’s a philosophical connection between writing and prompting AI. If you can describe something well (Wittgenstein’s “what we can say, we can will into existence”), you can code it, especially with some technical knowledge.
Detailed PRDs still win - Despite advances in AI, the biggest factor in one-shot success with AI coding is still a well-written product requirements document. Specificity directly correlates with better outcomes.
The bitter lesson applies to agents - Future AI development will require less handholding and detailed prompting. Natural language descriptions will be sufficient for personal use cases where scalability isn’t a concern.

Summary

Tina He (Twitter: @FakePixels) is a venture capitalist, former founder, and perpetual tinkerer who brings a unique perspective to AI development focused on joy and playfulness rather than pure productivity. Her career spans building a developer platform for smart contracts, selling her company to Coinbase, leading developer tools at Base (Coinbase’s layer 2 blockchain), and now investing at Pace Capital with a focus on developer tools and infrastructure.

In this session, Tina demonstrates a whimsical personal “operating system” website she built using Claude Code and Next.js - a vintage Mac OS-inspired web app containing multiple mini-applications including a “corporate speak translator,” music player, trading interface, image enhancer, and data labeling side hustle simulator. Rather than focusing on typical productivity use cases, she emphasizes how dramatically lowered costs of creation enable building purely for fun, laughter, and self-expression - fulfilling her childhood dream of creating joyful gaming experiences.

Her approach challenges the utilitarian framing of AI development, arguing that the 10% of usage dedicated to delight represents a meaningful shift in how we socialize and express ourselves creatively, soon to become as common as sharing other forms of content.

Main Topics

Background and Journey to AI Development

Tina traces her path from childhood tinkerer to AI developer, starting with RPG Game Maker (a niche Japanese tool) that first introduced her to programming. She describes herself as having three core threads: tinkering, writing, and investing.

[00:00:46.750] “I’ve always loved gaming. I think both as a storytelling medium, but also like over time kind of fell in love with just how you could design like a world where you can like honestly hang out with your friends and the social aspects of it.”

Her career combined these interests: product development, running a product studio in school, venture capital at Pace Capital (2019), founding a crypto company building developer platforms for smart contracts, and ultimately leading open source developer tools at Coinbase’s Base blockchain.

[00:02:56.790] “I feel like as a writer myself, and same with you, Dan, like, I think the technology of LLM is just like deeply philosophically interesting, where there’s this kind of interesting dichotomy of like the worst cell and the shape rotator discourse on Twitter specifically.”

The Philosophy of Vibe Coding

Tina reframes “vibe coding” as fundamentally about the vibe, not the coding - building what you want to build and having fun doing it.

[00:04:02.390] “It’s actually more about the vibe than the coding. It’s like, you are having fun building what you want to build. And you know, that’s actually the most important motivation.”

She notes that many of her best ideas come from jokes or practical needs like not wanting to pay for software and just building it herself instead. For writers and those with technical knowledge, the ability to describe something well translates directly into coding capability.

[00:03:20.470] “This is like, you know, actually a renaissance of worst cell’s power, like if you’re able to describe something, you can really well into existence.”

The Personal Operating System Project

Tina’s main demo showcases a vintage Mac OS-inspired web application containing multiple mini-apps:

Corporate Speak Translator: An app that takes frustrated workplace messages (like “You want me to do the work of a senior developer by paying me as a junior”) and translates them into professional corporate language. Uses Claude API for translations.

Music Player: Background music capability for the workspace.

Trading Interface: A demo trading app with 2X leverage options (not connected to real accounts, for demonstration only).

Image Enhancer: Uses Replicate AI to generate professional corporate backgrounds for photos.

Data Labeling Side Hustle: A playful take on the “AI economy” where you can do simple labeling tasks.

[00:14:00.430] “This app is my vision for the future of work. Like in the future, when you’re working, um, you can basically create your own customized workspace, um, for you to do all kinds of different random stuff.”

The entire project is live as a web app built with Next.js and deployed on Vercel.

Technical Stack and Tool Discovery

Tina emphasizes using Context7 as a key resource for understanding the current developer ecosystem:

Context7: An MCP (Model Context Protocol) server that provides agents with up-to-date documentation for popular tools and libraries. It functions as an “API for agents” to access external data and tools.

[00:15:51.890] “Tokens don’t lie. It’s like, basically, I love that. Oh, put it up on a banner. Let’s go. Tokens don’t lie.”

Context7 shows Next.js as the most popular framework, followed by others. The token allocation reveals what developers actually use versus marketing claims. She also mentions OpenRouter as another source for understanding model usage patterns, though accuracy depends on opt-in participation.

Her recommended stack: - Next.js: For web application framework - Vercel: For deployment - Tailwind CSS: For styling (#1 CSS framework) - Better Auth: For authentication (over Clerk, Auth0, Supabase, or Firebase) - Claude Code: Primary development tool - Replicate AI: For image generation/enhancement

Workflow and Development Practices

Plan Mode and PRDs: Tina emphasizes detailed product requirements documents as the key to success.

[00:20:01.160] “The biggest thing in, um, kind of creating really good, uh, good outcome. And like, usually with one shot success, it’s just like really well written PRDs.”

She notes that while a year ago extensive detailed specs were necessary, AI has improved but specificity still matters. The more detailed you are, the better the outcome.

Skills.md Files: Tina recommends using Anthropic’s new Skills feature - a file that describes how to precisely execute tasks, creating “soft in-context guard rails” for Claude Code.

[00:19:02.530] “If you’re like, uh, you’re a small startup, um, it’s, it’s like, it will be very efficient to store your UI system, like into like a, like a startup skill. Uh, and then every time you build a site, like they would just like refer to the style guide, um, of that skill file.”

Visual References: For UI design, she provides images of desired aesthetics (like vintage Mac OS screenshots) alongside text descriptions to achieve specific visual styles.

Development Tool: Primarily uses Claude Code for implementation.

Building for Joy vs. Productivity

Tina makes a deliberate case for the 10% of AI usage dedicated to fun and creativity:

[00:24:08.460] “I feel like those are going to get even like easier to do… There’s going to be less and less like handholding on the agent… In the future, we can really just describe what we want to do in natural language… so we stick to get pretty good results.”

[00:25:02.180] “The 10% of it, you know, is actually the meaningfully lower cost to create joy and fun. And, and I think that that’s why in the intro, I kind of like emphasize that gaming aspect. I feel like this is my childhood dream to be able to just like create fun experiences that make people laugh and make people have fun and, and, and enjoy, uh, kind of life a little bit more.”

She sees this as a new form of self-expression and socializing, predicting that people will commonly share custom-built apps with friends and loved ones, creating content “even like way more creative than this little experiment.”

[00:24:24.700] “Given the cost is literally lower than probably buying someone a cup of coffee, like definitely lower than buying someone a cup of coffee. Like it can just bring a smile on their face if you build an app for someone.”

The Future of No-Code AI Development

Tina acknowledges that tools are evolving to remove even more friction:

Claude Code: Still requires terminal, somewhat intimidating
Conductor: Creating better UI for coding
Lovable and Wabi: Companies removing the need to see any code at all

She envisions a future where creating and sharing apps becomes as common as other forms of content creation, with even more creative outputs than current experiments.

Actionable Details

Tools and Resources

Context7 (context7.io): MCP server for accessing up-to-date documentation. Shows most popular frameworks by token usage
OpenRouter: Tracks model usage by tokens across different use cases (finance, role play, etc.)
Claude Code: Primary development environment
Next.js: Web application framework
Vercel: Deployment platform
Tailwind CSS: Styling framework
Better Auth: Authentication library
Replicate AI: Image generation and enhancement
V0 (Vercel): Vibe coding tool
Monologue by Every: Voice-to-text tool
Whisperflow: Alternative voice-to-text (what Tina uses)

Development Workflow

Start with detailed PRDs: Write specific product requirements
Use visual references: Include images for desired UI/UX styles
Create Skills.md files: Store UI systems and style guides as skills for consistency
Leverage Context7: Check token usage to see what tools/frameworks are most popular
Use Claude Code: Primary development tool
Deploy on Vercel: For hosting web applications

Live Demo Debugging

When the corporate speak translator failed during the demo: - Issue was API credits running out (not code problems) - Solution: Refilled $20 in Anthropic credits - Debugging approach: Check error messages, paste into Claude Code

Technical Architecture

Web application built with Next.js
Deployed on Vercel
Uses Claude API for text transformations
Uses Replicate AI API for image operations
Vintage Mac OS UI aesthetic achieved through detailed prompting and reference images

Quotes Worth Saving

[00:15:51.890] “Tokens don’t lie.” - On why usage data from Context7 and OpenRouter reveals true adoption better than marketing

[00:03:20.470] “This is like, you know, actually a renaissance of worst cell’s power, like if you’re able to describe something, you can really well into existence.” - On the philosophical connection between writing and AI coding

[00:20:01.160] “The biggest thing in, um, kind of creating really good, uh, good outcome. And like, usually with one shot success, it’s just like really well written PRDs.” - On what actually makes AI coding succeed

[00:25:02.180] “The 10% of it, you know, is actually the meaningfully lower cost to create joy and fun… I feel like this is my childhood dream to be able to just like create fun experiences that make people laugh and make people have fun.” - On building for delight rather than pure productivity