From China to AI: Ryan Shuken on startup lessons and innovation
with Ryan Shuken, Startup Advisor & Investor
From China to AI: Ryan Shuken on startup lessons and innovation
Show Notes
Ryan Shuken is a startup advisor, accelerator director, and investor who spent over a decade in China helping build the startup ecosystem in Shanghai - running accelerators, directing investment programs at a major early-stage VC, and working hands-on with companies building cross-border into and out of China. He has since returned to the US, where he now works with early-stage founders on what it actually takes to launch, and runs the vibe coding workshop series at Denver's local AI meetup.
His current mission is direct: help founders not run into the same problems he has seen a thousand times. With vibe coding lowering the cost of software development to near zero, Ryan believes the entire playbook for starting and funding companies has changed - and the window to act on that shift is right now. You can find him at ryanchuan.com or connect on LinkedIn at Ryan Shuken.
The Two Things That Will Save Your Vibe Coding Project
Most vibe coding failures - the horror stories of 30,000 lines of code going sideways, losing weeks of work, shipping something broken - come down to two missing fundamentals: no project plan document and no version control. Ryan frames these as the difference between someone who vibe codes for a weekend and someone who actually ships something.
The PRD (Project Requirements Document). Before writing a single line of code, spend the first third of your total project time in pure planning. Use Grok, Gemini 2.5 Flash, or Claude to collaborate on a project requirements document - a markdown file that defines the app's goals, functions, design direction, and user stories in plain language. This becomes the north star of the project, saved directly in your codebase. The key instruction: tell your IDE (Cursor, Windsurf) to always reference the PRD before doing anything. This prevents the AI from hallucinating new features, adding scope, or rebuilding parts of the app you didn't ask it to touch.
Git version control. The AI is not saving your work the way you think it is. Every prompt rewrites the same files - it doesn't create backups. Git is free, takes one setup session, and creates discrete save points you can roll back to. Think of it like save states in a video game: before you start a new feature, create a save point. If the feature goes wrong, you restore the save. Rule of thumb: commit early and often. You can prompt your AI to handle all of this for you once you understand what it's doing.
How AI Just Flipped the Startup Investment Model
When Ryan was running a VC accelerator and making $1M early-stage investments in China, roughly $800,000 of every investment - 80% - went to development: front-end, back-end, middleware, security, project planning, and developer salaries. That was the cost of getting a working product built. The other 20% was everything else: team, marketing, sales.
That math is now inverted. Development today costs a fraction of what it did - $10,000 to $100,000 at most to get something real out the door, with two people using AI tools outputting what used to require an entire engineering team. The 80% that used to go to development now goes to market - building the team, running marketing (which AI also assists with), generating sales. The compounding effect: Ryan genuinely believes a fully vibe-coded unicorn - a billion-dollar company built by non-technical founders - is imminent. Not hypothetical. Near-term.
This also breaks the VC model as it was designed in 1980s Silicon Valley. Small checks now do what large checks used to. A few hundred thousand dollars can fund and accelerate a company that previously required $3-5M just to get to working software. The firms fast enough to restructure their fund documents and LP commitments around this reality will have a structural advantage; the ones locked into old PPMs will miss the wave.
China's Startup Ecosystem: What the West Gets Wrong
Ryan spent over a decade watching the Chinese startup ecosystem evolve from the inside, and his read on the IP question is more nuanced than the standard Western narrative. Chinese companies are not ignoring IP out of disrespect - they are moving as fast as possible to service a market where the competition is so intense that speed matters more than protection. The domestic competition Ryan witnessed in China makes American startup competition look mild. Companies were iterating, copying, and improving on each other at a velocity that has no US equivalent.
The IP transparency point is also important: when foreign companies go to China to manufacture or build, they sign agreements that clearly state their IP is not protected under Chinese law. It is disclosed up front. The companies that complain afterward agreed to those terms in exchange for access to factories, labor costs, and infrastructure. Ryan's view: that's a disclosed cost of doing business there, not a hidden violation.
More importantly for founders today: if you still think of China as “copycat China,” Ryan says you are eight to ten years behind. China has been in innovation mode for years - electric vehicles, delivery logistics, e-bikes, restaurant technology, drone applications. Much of what the West considers cutting-edge has already been iterated on in China. The drone swarms over cities, the robotics in construction, the speed of infrastructure builds - these are not imitation; they are original innovation at scale.
The Future of Human-AI Interfaces: Ambient, Invisible, Attention-Based
Ryan's vision for where AI goes next in terms of human interface is not more apps - it is the disappearance of apps. The fundamental problem with technology interfaces until now is that they all operate on an interrupt model: notifications, alerts, alarms, badges. Every interaction requires the human to stop, look, respond. Steve Jobs hated this design pattern. It is anti-human by design.
AI is the first technology genuinely smart enough to reverse this. Instead of demanding your attention, it can anticipate what you need and handle it without being asked. Ryan gives the example of a morning meditation that transitions into an AI-delivered briefing - personalized, researched, delivered in a soothing voice, with nothing for the user to do except listen. Or ambient attention-tracking: if the AI notices you looking toward the kitchen, it gently adjusts the lighting over there. Not because you asked. Because it understood.
The infrastructure making this possible is MCP (Model Context Protocol) servers - the layer that connects AI to actual tools and services. Any founder building a product today can publish it as an MCP microservice, making it accessible to any AI system that a user runs. Instead of building an app that requires a download and account creation, you build a capability that integrates into whatever AI interface the user already has. Ryan frames this as the architectural shift that makes truly frictionless AI integration possible.
One-Shot Coding vs. PRD-Based Vibe Coding
There are two distinct modes of vibe coding and they serve different purposes. One-shot coding - using tools like same.dev or Lovable to generate a working app from a single prompt - is essentially throwing a dart and seeing what the AI hits. It's fast, fun, and occasionally produces something usable. It is not how you build something that ships to real users and makes money.
PRD-based vibe coding is the structured approach: plan first (10-12 hours for a weekend project), document everything, then build feature by feature against that document. The PRD lives in the codebase, the AI is instructed to always reference it, and every new feature gets its own git save point before you start. This is how you get to production. Ryan's rule: if you want to mess around and see what AI generates, one-shot. If you want to actually ship something and make money, PRD first, always.
Tools & Resources Mentioned
- Cursor - Ryan's top IDE recommendation for vibe coding; supports global and local AI instructions including PRD referencing
- Windsurf - mentioned alongside Cursor as an IDE improving at managing code saves and version states
- Grok (xAI) - recommended for collaborative PRD creation and project planning before building
- Gemini 2.5 Flash/Pro (Google) - recommended alongside Grok for PRD planning; Ryan calls it very good at collaborative planning
- GitHub / Git - free version control; essential for vibe coding save states; can be managed entirely through AI prompts once set up
- Lovable / same.dev - one-shot coding tools; good for exploration and rapid prototyping, not production builds
- Bolt - mentioned in context of the Bolt hackathon; free access available at the time of recording
- MCP (Model Context Protocol) - the server layer allowing AI to access tools and microservices; Ryan describes it as the infrastructure backbone of future ambient AI interfaces
- Zero to One - Peter Thiel - referenced as a touchstone for early-stage founder philosophy; Ryan identifies strongly with the zero-to-one moment
- Google Glass (augmented reality) - Ryan worked on AR sign translation using Google Glass before the technology was mainstream
Frameworks
PRD First, Always
Before writing a single line of code, spend a third of your total project time building a Project Requirements Document using a planning AI (Grok, Gemini, Claude). The PRD defines goals, functionality, design, and user stories in plain language and gets saved as a markdown file inside your codebase. You then instruct your coding IDE to always reference this file before making any changes. The PRD prevents scope creep, hallucination drift, and the AI silently rewriting parts of the app you didn't ask it to touch. Without it, you are building on sand.
Vote Early and Often (Git Commits)
Version control is not optional for vibe coding - it is the difference between a project that ships and one that collapses. Git creates discrete save points: before each new feature, create a commit. If the feature breaks something, roll back to the save. Think of it like save states in a video game with branching paths - you can explore one direction, save, try another. Your AI can handle all the actual git commands once you understand what it is doing. The mantra: commit early and often.
The 80% Inversion
Pre-AI, roughly 80% of an early-stage investment went to development - engineers, infrastructure, project management. 20% remained for everything else. Post-AI, that ratio inverts: development now costs a fraction of what it used to, and 80% of resources go to market - team building, marketing, sales. The practical implication is not just that startups are cheaper to build; it is that the bottleneck has moved from technical execution to go-to-market strategy. The founders who understand this will raise smarter, hire differently, and spend their first dollars in the right places.
MCP as Product Distribution Strategy
Instead of building a standalone app that requires users to download, sign up, and learn a new interface, founders can build their product as an MCP server - a microservice that any AI can call. The user runs one AI on their phone or home, and your product integrates into it. The user doesn't download your app; they already have it because it's part of their AI's capabilities. This is the emerging distribution model for ambient AI products, and Ryan frames it as the right way to think about product architecture for founders building right now.
Innovation Beats Protection
Ryan's lesson from over a decade watching the Chinese startup ecosystem: the companies that tried to protect their IP lost. The companies that moved fast, iterated faster, and kept innovating won. China internalized this - going from copycat to innovator in under a decade - because the domestic competition was so intense that speed was the only real moat. The takeaway for Western founders: your best IP protection is never a patent, it is always moving faster than anyone who might copy you.
FAQ
What is a PRD and why does vibe coding fail without one?
A PRD (Project Requirements Document) is a plain-language document that defines what you are building: its goals, features, design direction, and user stories. In vibe coding, the AI will hallucinate, drift, add scope, and silently rewrite parts of your app unless it has a fixed reference point to work from. The PRD is that reference point. It gets saved as a markdown file inside your codebase, and your IDE is instructed to reference it before every change. Without it, every prompt is a fresh roll of the dice - sometimes additive, sometimes destructive. With it, the AI stays on track through the entire build. Ryan recommends spending 10-12 hours on PRD creation before writing a single line of code, using planning-focused models like Grok or Gemini to build it collaboratively.
What is Git and why does it matter for vibe coding?
Git is a version control system - a way of creating discrete save points in your code, branching to try new features, and rolling back if something breaks. It is free and available through GitHub. Without Git, every AI prompt rewrites the same files with no history and no rollback. If the AI makes a change that breaks something three prompts later, you may not even notice until it is too late, and you have no way to recover the working version. With Git, you commit before every new feature, creating a restore point. Your AI can manage all Git commands for you - you just need to set it up once and give it instructions to commit regularly.
How has AI changed the economics of building a startup?
Historically, 80% of an early-stage investment went to development - engineers, infrastructure, and the overhead of managing them. Today that number has collapsed. A working product can be built for $10,000-$100,000 with a small team using AI tools, compared to $800,000+ before. The 80% that used to go to development now goes to market: team building, marketing, sales, and growth. This means founders can reach product-market fit on dramatically less capital, VCs can write smaller checks that go further, and the barrier to building something real has never been lower.
What is one-shot coding and when should you use it?
One-shot coding is generating an entire app from a single prompt using tools like same.dev, Lovable, or Bolt. The AI takes your description and outputs a working (or near-working) application in one pass. It is the fastest path to seeing something on screen and is genuinely useful for exploration, prototyping, and learning what is possible. It is not the right approach when you want to build something that ships to real users, scales, or makes money. For production builds, use the PRD-based approach: plan first, build incrementally, commit often.
What does Ryan think the future of AI interfaces looks like?
Ryan's vision is ambient and invisible AI - technology that handles things without demanding your attention rather than interrupting you with alerts and notifications. The current interrupt-based interface model (notifications, badges, alarms) is bad design that creates anxiety and stress. AI is the first technology smart enough to reverse this: it can anticipate needs, take actions without being asked, and communicate in a way that feels like talking to someone rather than operating a machine. The infrastructure enabling this is MCP (Model Context Protocol) servers, which let any AI access tools and services as microservices - meaning founders can build products that integrate into whatever AI a user already has, rather than requiring new app downloads.
What is Ryan's actual take on China's startup ecosystem and IP?
Ryan spent over a decade inside China's startup ecosystem and his read is more nuanced than the Western narrative. On IP: when foreign companies manufacture or build in China, they sign agreements explicitly stating their IP is not protected under Chinese law. It is disclosed upfront. The 'they stole my IP' complaints often come from companies that agreed to those terms in exchange for access to cheap manufacturing. On innovation: if you still think of China as copycat China, you are 8-10 years behind. China has been in a full innovation mode for years - EVs, delivery logistics, drone technology, robotics, restaurant tech. The domestic competition inside China is so intense that moving fast and innovating continuously is the only real competitive strategy, and they internalized that lesson before most Western markets did.