The Gap That Didn't Close

Richie McIlroy, "Just Start Building":

The next generation of great software is not going to come from the companies with the biggest engineering teams or the deepest pockets. It's going to come from individuals and small teams. People with taste, with opinions, with a problem they actually care about solving, who happen to have access to tools that didn't exist two years ago. The playing field hasn't just levelled. It's flipped. A single person with conviction and the right tools can now outship a team of fifty who are building by committee.

There are two posts stitched together here, and one of them is good.

The good post is about Resistance — the Steven Pressfield kind, the force that makes you reorganise your Notion board instead of writing the first line of code. That post is right. The psychological barrier to starting has always been the real killer, and AI tools have made the first hour of a project dramatically more productive. Starting is easier than it has ever been. That matters.

The other post is the one where "starting is easier" quietly becomes "shipping is easier" and then "competing is easier" and then a single person outships a team of fifty. That post is doing a lot of work to flatter its audience.

Here's what actually closed: the gap between an idea and a demo. You can describe something to Claude Code or Cursor and have a working prototype in an evening. That's real and it's transformative for exploring ideas.

Here's what didn't close: the gap between a demo and a product. Auth. Error handling. The edge case that only appears when a real user does the one thing you never considered. Security. Accessibility. Infrastructure that stays up when someone links to it on Hacker News. Support, when it doesn't. That's where the teams of fifty earn their money, and no amount of AI-assisted scaffolding has changed that.

The most revealing line is the claim that "two focused hours at night with an AI agent can produce what used to take two weeks." For boilerplate, sure. For wiring up a standard CRUD app, maybe. For anything involving real architectural decisions or domain complexity — the kind of decisions that are the product — the compression ratio is nowhere near that. And the people most likely to believe the two-hours claim are the ones least equipped to spot when the agent has made choices they'll regret in month three.

The suggested workflow is actually sensible: start with a brutally simple description, let the AI interrogate the idea, build a living spec, review everything the agent produces, treat it like a fast junior developer who needs supervision. Good advice. But notice what that workflow requires to work well: enough technical judgement to review the output and enough product sense to know when "done" means done. That's not "anyone with an idea and a terminal." That's a experienced developer who's also a good product thinker. A considerably smaller group.

Pressfield's original point about Resistance was that you have to sit down and do the hard work yourself — that there are no shortcuts, and the shortcuts are what Resistance sells you. There's an irony in repackaging that idea as "open the terminal and tell an AI to build it for you."

The best thing you can do with this post is take the first half seriously and hold the second half at arm's length. Open the terminal. Describe the dumbest version of your idea. Start building tonight instead of Monday. All true. Just don't mistake the prototype for the product, or the starting line for the finish.