I love building things. With AI I can now build almost anything I imagine, and that’s exactly the danger.
At my last startup, my team knew how to contain me. Every time I opened a meeting with “I was thinking…”, the engineers and product team braced, because a new feature was coming and they knew exactly what it would cost. That friction was a feature. AI removed it.
Not long ago I built a working demo of a product idea in a few days, something that would have taken a funded team months just a few years back. It was fun. Every day I caught myself wanting to add one more screen, one more flow, one more small feature that quietly became part of the product. Nobody prices the build anymore, so the restraint has to come from me.
It’s not just me. The same pitch keeps showing up lately, mostly from new builders, sometimes from experienced ones. It starts well. A sharp wedge, a real problem, a founder who clearly lived the pain. Then the demo begins, and the wedge has company. There is a Yelp for the wedge. A TikTok clip for the wedge. Sometimes a dating app for the wedge. All of it built. All of it working. Not a single customer asked for any of it.
Five years ago that pitch was impossible. Building three products took three teams and a real budget. Today one builder with AI tools and a few weekends can ship all of it. These builders are talented and fast, and most never worked through a time when one feature cost three months and two engineers. When building feels free, the discipline of not building never gets a chance to develop.
That discipline is what this piece is about.
The MVP solved a different problem
Frank Robinson coined the term Minimum Viable Product in 2001, and Eric Ries made it mainstream through The Lean Startup, a classic that never gets old: get the maximum validated learning with the least effort. That idea created a generation of discipline. Don’t disappear for two years polishing the perfect product. Don’t confuse your roadmap with customer truth. Ship the smallest thing that teaches you something.
It was the right lesson for an era when building was slow and expensive. In that world, startups died from building too slowly. In this one, I believe many will die from building too much, too fast. Not because the MVP is wrong, but because building has changed underneath it.
Protect the instinct, not the idea
Mark Pincus has a useful frame for the step before the MVP. In his excellent new book Life at the Speed of Play, he separates instincts from ideas, and in a conversation with Shane Parrish on The Knowledge Project he put it bluntly: “our instincts are almost always right and our ideas are usually wrong.”
A founder’s instinct is often correct. Something is broken, a behavior is shifting, a new technology makes something newly possible. The first product idea attached to that instinct is usually wrong, and this is where founders get trapped: they protect the idea instead of protecting the instinct. Pincus pushes teams toward what he calls a minimum idea state, something rough enough to vibe code and put in front of people, instead of polishing an MVP around an idea nobody has validated. In the same conversation he goes further: too many teams waste time getting a single MVP to market, when what they need is a “failure machine” at the top of the funnel, testing many ideas cheaply.
None of this replaces the oldest rule in the book. YC’s motto is still make something people want, and Paul Graham has spent years telling founders to start from real problems, ideally their own, instead of inventing ideas in the abstract. AI changes none of that. It can help you build faster, but it can’t make people care. It can generate features all day, but it can’t tell you which pain is worth owning.
A prompt is not a product
For AI products specifically, there’s a question I now ask before anything else: can this workflow become promptable, repeatable, and trusted?
Not can the model do it once. Anyone can get a model to do something impressive once. The question is whether the product can do it again and again, inside a workflow that matters, with enough context, memory, permissions, and reliability that the user keeps coming back after the novelty fades. A prompt isn’t a product. A chatbot isn’t a product. A great demo isn’t a company.
Every AI startup should be able to say this in one sentence: when the user expresses this intent, the product produces this outcome, and the outcome matters enough that the user returns. If that sentence is clear, specific, and painful, there may be a company. If it sounds broad or magical, there’s probably just a demo.
The new failure mode: prompt creep
The old software world had feature creep. This one has something faster and more seductive. Call it prompt creep (a name that, fittingly, came out of drafting this essay with an AI).
Prompt creep is what happens when every “AI can do that” becomes a feature. You start with a sharp wedge. Then the product also writes emails. Then it summarizes meetings. Then dashboards, contracts, web search, reports, agents. Each addition felt nearly free to build, and somewhere along the way the product lost the plot. The Yelp-for-the-wedge pitch is the extreme case, but milder versions are everywhere.
The feature is cheap, but it’s never free. Every new capability adds surface area: more edge cases, more permissions, more evals, more bugs, more support burden, more ways to fail, and more confusion about what the product is. AI lowers the cost of building. It does not lower the cost of complexity.
The questions that matter now
Investing and building keep bringing me back to the same questions.
Before the MVP: what is the smallest version that proves people care?
Before the AI product: what is the one workflow this company should own?
Before every new feature: does this strengthen the core, or is it prompt creep?
Before scaling: is the market pulling us into this, or are we pushing because it is fun to build?
That last one matters more than it sounds. The market pulling you into adjacent workflows is expansion. The team pushing features because they are exciting to build is distraction dressed as progress.
On the investing side, founders still come first. Founder quality and founder-market fit are the filter before anything in this piece, and that deserves its own post. Assume the founder is already vetted. After that, the signal is no longer whether the team can build. Almost everyone can build now. The signal is whether they can explain their core workflow in one clear sentence, and whether they have the discipline to stay inside it until the market pulls them out. The best AI startups often look narrow at first. Narrow is fine. Narrow is how most great companies start. The mistake isn’t being narrow. It’s being shallow.
Restraint is the new advantage
AI is making small teams more powerful. Everyone can see that. What’s less obvious is that it makes restraint more valuable. When anyone can build almost anything, the winners will be the founders with the discipline to build only what matters. They will move fast, but not randomly. They will test early, but they won’t fall in love with their own demos. Most of all, they will know what not to build.
None of this is an argument for going slow. It’s an argument for going fast in the right order. In Blitzscaling, Reid Hoffman makes a case that still holds: once you have product-market fit, you accelerate hard and prioritize speed over efficiency. What AI changes is how fast you can move on the product side once you get there, turning months of building into days. The mistake is blitzscaling before the market has told you what to build. Restraint is what you practice until you earn the right to floor it.
The products that endure won’t be the ones with the most AI surface area. They will be the ones with the clearest core workflow.
This essay is proof of its own point, I’ll admit. I am an avid reader, not a natural writer, and AI helped me shape this piece. It raised the ceiling on what I could make, and it tempted me past it: my first draft ran more than twice this length, with four named frameworks and a six-step sequence. Cutting it hurt. Even writing about overbuilding, I overbuilt. The trap doesn’t care how experienced you are. So keep building. If you’re building for fun, overbuild; playing is how you learn what this era makes possible. When you’re building for customers, make sure it’s something people want, not something you build because you can.
The MVP saved a generation of founders from overbuilding before learning. This era needs the same discipline, probably more of it. Because the danger is no longer building the wrong thing slowly. The danger is building the wrong thing beautifully, quickly, and with total conviction.
That may be the most expensive trap of all.
Hector Hulian is the Founder & GP of Seedradar Ventures, investing in early-stage founders primarily across North America.
Further reading
At Seedradar Ventures, we back relentless founders at the earliest stage, investing in teams solving big problems with solutions that deliver a 10–20x greater impact.
Our syndicate offers investors access to a curated pipeline of early-stage startups, sourced through our network.
🔎 Powered by Seedradar Ventures
Disclosure: Seedradar Ventures and/or its general partner may hold investments in some of the companies mentioned in this newsletter. This content is provided for informational purposes only and should not be considered investment advice.
