The 48-Hour Prototype (and the Monday Token Invoice)
We spent the last five years making software development thoroughly miserable. We wrapped it in endless YAML files, buried it under heavy Docker containers, and metrics-driven-developed it until it tasted like lukewarm tap water. We turned the garage into a cubicle farm and called it progress.
But the joke is on them.
Because sometimes, on a rainy Saturday, after the quarterly roadmaps and compliance decks and productivity dashboards have all slithered back into their folders, you sit down with a blank file and a machine that does not sleep, and forty-eight hours later there it is: a small, functional, strangely beautiful tool that did not need permission to exist.
It feels like theft. The clean, intentional kind. Not stealing from people, but stealing back a little agency from the framework-industrial complex.
That is the thrill running through Simon Willison’s note on Thariq Shihipar’s “Unreasonable Effectiveness of HTML”. A single HTML file can still be a complete little machine. Not a platform. Not a transformation initiative. Just a document with teeth: layout, diagrams, interaction, explanation, utility. A bicycle for the mind with fresh grease on the chain.
This is where the joy gets back in through the side door. The models are not thinking in any human, accountable sense. They are text-prediction engines with a gift for boilerplate, syntax, and the kind of tedious connective tissue that used to drain the color out of a weekend. We know the risks. We use them anyway.
Round The Code’s account of building an AI tool for .NET developers in 2 days shows the bargain. Useful scaffolding appeared fast, but not flawlessly. The machine generated most of the code, then handed back compile errors, mismatched package assumptions, and the usual haunted furniture of modern development. Which is exactly the point. The tool accelerates the work. It does not absolve you from understanding it.
Model Context Protocol belongs in that same category. It is mostly plumbing, which is often better than magic: a way for assistants to discover and call tools you define. Less oracle and more patch cable. Examples such as MCP Server for .NET help show us the shape of it.
This isn’t about escaping reality; it’s about navigating it without surrendering judgement. Sean Goedecke’s How I use LLMs in 2026 puts it plainly: use the machine to accelerate your intent, not replace your judgement. The moment you outsource comprehension, the hangover starts. The magic and the rot are part of the same pipeline.
While solo hackers are rediscovering the old joy of making small things that work, corporate engineering floors are busy manufacturing a stranger crisis. We are training developers to treat software like a séance, knocking on the table for clean Python and panicking when the spirit returns a runtime error.
If you want to watch the industry rebrand educated guessing as an engineering style, The New Stack’s dispatch on the ACM Vibe Coding AI Agent is a useful specimen. The follow-up warning is even less comforting: we may be raising an AI generation who can’t debug. Not because they are lazy, but because the toolchain keeps teaching them that code is something you summon, not something you build and understand.
The industry calls this velocity. Executives point to productivity booms. We see the telemetry differently. Senior developers are left holding the tab for a mountain of cognitive debt, compounded daily at usurious rates and paid in eye strain, cold coffee, and the quiet horror of debugging code that nobody actually read. If you’re cleaning up automated spaghetti code right now, the O’Reilly Radar analysis on Burnout and Cognitive Debt will feel less like analysis and more like a workplace injury report.
Addy Osmani’s Don’t Outsource Learning is a clean version of the rule: acceleration is useful; intellectual surrender is not.
The outlook is compromised. The corporate noise is deafening. But the terminal is still open, and that still matters.
The only honest response is a stubborn refusal to turn off your own brain. Use the machine. Doubt the machine. Make it show its work. Keep one hand on the keyboard and the other on the eject button.