Interesting observation I’ve noticed from Anthropic:
They want to dominate a market they're actively making more competitive (à la Jevons paradox).
Anthropic is reportedly building a Lovable competitor inside Claude. Meaning they're racing to make front-end vibe-coding faster, easier, and cheaper.
At the same time, the same labs are building MCP, Sonnet, Opus, agentic frameworks - tools that actively reduce how much front-end software anyone will need to build
I loved the gumption of the Rabbit r1 team and wanted it to succeed, but it’s a textbook example of failing to meet a fundamental product principle:
The best products will meet you where you are.
What that means is the product comes to you, to the device you already use, and it improves your workflow.
R1 was supposed to be ambient AI in your pocket; order Uber, play Spotify, check flights. But the problem was, it asked you to carry a new device that wasn’t an improvement on the iPhone.
For any tech to be successful, the UI has to be thought about from first principles. It has to add convenience, it has to be simple, and it shouldn’t disrupt the current workflow.
If you want users to change their “routine,” the product needs to be a 10x improvement. And since Rabbit R1 didn’t meet that bar, there was no incentive to change.
This is exactly why MCPs are becoming so powerful. They do not ask you to adopt a new tool. They interlink the tools you already use. Your Slack, your calendar, your CRM.
The intelligence sits on top of the existing ecosystem.
(and why I’m so bullish on the future of software & hardware in the era of MCPs).
Shopify has an internal adage I think about often:
99% of process doesn't need to exist as long as you have a high-agency, high-performing person in the role.
Especially when you’re a service-based business.
The remaining 1% that does, you cannot skip - legal, onboarding, or any non-negotiable processes for compliance or safety.
But the rest, you cannot substitute with documentation, no matter how thorough.
In the 90s, Larry Ellison used to tell Oracle customers:
”You are better off with an 80% solution installed and working in 6 months than fantasizing about a 100% solution that you might finish in two years after you write lots of custom code.”
Software had clear and rigid constraints: time, money, and talent.
Since then, we’ve made gradual improvements to the so called “customization allowance” Ellison refers to in the video when it comes to building software.
And today? We have finally reached the point where 100% solutions are viable. Thanks to MCPs and agentic workflows there is no longer an excuse NOT to deliver exactly what customers want.
This is what excites me more than anything…helping enterprise companies execute on their vision and build out that last 20%.
“Wtf is an MCP?” - if you’re building a product you’ve probably had that question in the last month.
An Model Context Protocol (MCP) is effectively a communicable layer on top of an application that an LLM or an agent can access to understand what a customer wants to do.
In practice, it means instead of opening your meeting recorder software and clicking around to find a transcript, you type "send me yesterday's meeting next steps" and an agent pulls the transcript, the recording, and drops it in Slack.
You say what you want; the MCP communicates to the agent, and the agent handles it.
I think the Lean Startup has done enormous harm to technology community.
The Lean Startup told us to build scrappy, think MVP, build for just a few users. What we got was a lot of ugly software that can’t scale.
This is one of the most expensive mistakes I have observed founders make who had incredible insight.
Better way (what we do at Rare Days): when scoping a v1, every architectural choice gets stress-tested against the 18-month version of the product.
Some things still slip through, but it’s a helpful filter to future pace product decisions and catch 9/10 future redesigns.
Over the past six years, we've built product and rebranded for F500s with no real process.
May sound counterintuitive but it’s truly the only way to do it.
So when a customer asks how we work, here's what we do…
1// Understand the customer of our customer.
There's always a difference between what the customer says they want and the complex reality of our target user:
a) Who they are
b) Where they are when they use the product
c) What their routines look like
d) What they're constrained by
Customer obsession is not a sometimes-obsession at Rare Days, it is a true core value. Each customer is unique, each product is unique, and the only consistency in our process is that we start from scratch every time so each one stays unique.
We anchor to the psychographic and demographic profile of the end user, and managing expectations against what's possible to build is part of this step too.
2// Write out all the “jobs-to-be-done.” Then ask, for each one: does this need a screen, or can it be done by an agent?
This question has a very different answer in 2026 vs 2022 - because most of what gets built as a UI today would be better served by a chat interface or an agent doing the job invisibly.
The screens that survive are the ones where the user genuinely needs to see information at a glance i.e. calendars, photos, and any dashboards that require visualization.
Almost everything else can become a single chat input.
3/ Get the service design textually clear (a machine-readable document an agent can execute against).
Vibes-based specs create hallucinations in agents and mistakes in humans.
If you want to build an agent that executes a multi-step task across multiple systems, the instruction has to be precise and structured. This is the part of our process we don't compromise on - and it's worth the time we put into it.
Otherwise, that's pretty much it.
Everything else we adjust by team, by project, by moment.
The reason this works and a 200-step SOP doesn't comes down to one word you've probably heard floating around a lot:
Taste.
It's becoming a buzzword because few people actually know what it looks like.
The way I think about it: taste is pattern recognition. It's understanding what really good products do - what makes them delightful, what makes them work - and titrating those insights into the next thing you build.
It comes from being a student of software, a student of business.
Enough repetition interacting with the best businesses, and you start to develop a sense of what makes them great. Then you try to export that sense into the next thing.
Beautiful design is part of it, but the bigger part is choosing what to build and how to build it in the first place.
You can't write that down.
There's no SOP for "what's worth building" or "what good looks like" or "what to cut." That comes from the people in the room.
Which is why we hire for taste first.
If you're hiring an agency and they hand you a six-stage waterfall checklist as the answer to "what's your process," that's almost always a sign they don't have the people who can execute without one.
Vibe coded tools enable you to make software for the pre 2026 paradigm.
Some of that will be great no doubt.
But it is distracting for true innovation because they are indexed on old paradigms of software.
Descript is one of the cleanest examples of AI integration I’ve seen.
Instead of the traditional clipping UI, you're editing through a chatbox. It’s more intuitive and part of a broader trend we’re seeing - software is becoming something you instruct rather than something you manually go into and operate.
I train all my ~60 employees to be fundamentally anti-SOP at my engineering & design studio.
Here’s why…
Process is what we collectively throw at under-performers instead of replacing them.
I'll explain what I mean by this.
I once worked at a company that paid three full-time people to write SOPs. I don't think I ever saw anyone open one of them. That team existed because the company had hired people who couldn't do the work without instructions.
Rather than fix the hiring, the company papered over it with documentation. Every time a customer issue came up, the response was the same: we need to document this better. But we never REALLY did - we just generated more SOPs that nobody read.
Another important consideration (especially in 2026) is that process is inherently slow. It’s inherently going to change, so any 40-page SOP you write today is going to be partially obsolete in six months.
I’ll caveat this by saying I'm running a ~60 person company.
So of course I’m sympathetic to the argument that scale eventually demands process. I just think most companies reach for it too early and use it as a substitute for hiring well, rather than a complement to it.
The right person can adapt to a new situation in ways an SOP never can.
The wrong person needs an SOP to do anything at all - and even then, it usually doesn't work.
tl;dr
Process is the wrong solution to the most common problem it gets used for.
That's why I'm fundamentally anti-SOP.
Building software in 2026 reminds me of the Model T era.
When people got into the first Model T, nobody was thinking about seatbelts, headrests, or gloveboxes. We didn't know what cars wanted to be yet.
That's where we are with AI products right now. Most of what we're building today will look obviously wrong in five years - and we don't yet have the vocabulary to describe what's missing.