Most posts in this series are about specific, learnable behaviours. Write the docs first. Verify before you claim. Build error handling before the demo feature. Rules with clear edges that you can install and follow.
This one is different. This one is about the larger shape of six months of work — the strategic wrong turns, the moment things clicked, and what I actually believe now that I didn't believe at the start. It's messier than the other posts, because it's about thinking changing, which never happens cleanly.
The AI was present for all of it. Not as the cause of the wrong turns, but as the thing that made them move faster and get corrected faster — which is either better or worse than going slowly in the wrong direction, depending on how you look at it.
The original bet
I started this stretch building standalone Wix widgets. A cinematic hero animation for Wix sites. A phone-to-site editing tool. A quiz and knowledge-check engine for industry professionals. Each app had its own App Market listing, its own positioning, its own onboarding flow, its own small surface area of revenue possibility.
The logic felt sound: build useful, self-contained things, distribute them through the platform's own marketplace, let the platform do the discovery work. The Wix App Market has millions of sites to reach. Why wouldn't that work?
What I was building wasn't wrong, exactly. The apps were real. Some of them were genuinely good. The problem was the unit of product I was selling. A standalone hero widget — even a beautiful one — is a thin reason to install an app. It competes with a hundred similar tools, it doesn't justify a recurring fee on its own, and it can't survive the Headless era, where the developers who actually want to use advanced tooling are assembling sites from components, not browsing a marketplace.
I spent a significant amount of time building things that were individually well- made and collectively missing the point.
The lesson from the article
In April 2026 I read something that stuck: the argument that AI agents will increasingly operate services via APIs and CLIs, bypassing visual UIs entirely. The beautiful dashboard isn't where the value lives in an agent-mediated world — the API surface is.
I'd been doing the opposite. I was over-investing in dashboards and widget UI. I was building onboarding flows for a user who would navigate the product once, by hand, and then mostly not touch it. I wasn't building the kind of API surface that an agent could call on behalf of that user — the thing that would make the product composable, automatable, durable.
This didn't immediately change what I was building, but it changed how I thought about what I was building toward. A quiz engine with a clean API is infrastructure. The same quiz engine as a standalone widget is an App Market listing competing with fifty others.
The reframe was slow and it wasn't complete until the end of May.
The day it clicked
In a single session at the end of May 2026, a year's worth of product decisions rearranged themselves into something that made sense.
The thesis that locked: standalone widgets can't survive the Headless era as standalone products. The saleable thing — the real product unit — is an orchestration layer. Multiple DV apps bundled into a Wix Headless site as one coherent vertical: a creator portfolio, an AI consultant presence, a service business platform. The widget is an ingredient. The skill — the vertical bundle — is the product.
But a Headless site without a way for the owner to maintain it is a dev-locked artefact. Two surfaces are non-negotiable: a Wix Dashboard integration so the owner can edit content without touching code, and a Claude-driven maintenance layer so they can make updates through natural language rather than engineering tickets. Without those two things, the bundle doesn't scale — you've just done a bespoke build that can't be sold twice.
That architecture — audit → skill → ongoing maintenance — also produced a customer journey that hadn't existed before. An AI-era readiness audit finds the gaps. A Headless Skill fills them. Ongoing Claude-powered maintenance keeps them filled. Three products that sell each other instead of three products competing for the same marketing budget.
Nine apps got parked that day. Redundant, overlapping things that were each individually reasonable and collectively a distraction. The portfolio went from eleven App Market listings to six. Not because the work had been wasted — some of those parked apps became candidate ingredients for the new skills — but because the strategy had changed and the portfolio needed to change with it.
What the AI contributed
Claude was in the room for all of this. The session where the thesis locked wasn't me arriving with a finished idea and dictating it — it was a working session where the thinking developed in conversation. Stress-testing the thesis, identifying the three questions it couldn't yet answer, deciding which products to park and why.
The speed is the meaningful thing. A strategic pivot that might have taken three months of slow realisation compressed into a few weeks of working sessions because every idea could be tested against a counterargument immediately, every portfolio decision could be analysed against the strategy, every "but what about X" could get a real answer rather than "I'll think about it."
That's not the same as the AI being right. It wasn't always right. But it was always engaged, and always willing to hold the pressure on an idea until it either stood up or didn't. That's the collaboration that actually changes thinking.
The honest verdict
The thesis I landed on in May 2026 is a six-month bet, not a two-year plan. It needs three things to prove itself: one paid Skill installation, one audit that converted into a Skill engagement, one Build Partner who shipped a Skill they own. None of those have happened yet. The strategy is sound on paper; it needs to be sound in practice, which is different.
What I know is that I spent the first part of this period building in the wrong direction — not because I was careless, but because I hadn't yet seen the shape of the thing I was trying to build. The iteration that produced the right direction was faster and cheaper than it would have been without an AI that could hold the complexity of the whole portfolio in a single session and help interrogate it honestly.
Getting the strategy wrong is normal. Getting it wrong for two years instead of six months is the expensive version.