Twelve hours in Claude Code last week. Black screen, blinking cursor. My VC friends take one look and close their laptops. They don't get it.
Last week, my Wi-Fi was dying. Mac bar showing full bars. Classic. Told Claude Code to find what's actually wrong.
Sure, I worked at Meter, a networking hardware company. Know all the commands. But watching CC instantly surface twelve networks battling on channel 6? The interference patterns, signal-to-noise ratios, all of it appearing faster than I could've typed the first command? Different game entirely. Switched to channel 11 on my router. Fixed.
This isn't about not knowing commands. It's about not needing them.
Same session but now I'm building: poster for our Overnight Success event (pulled design patterns from my website, Monotype font and all), investment memo template, spreadsheet analysis Excel won't even attempt.
Commands typed? Zero. Just explained what I wanted and watched twelve hours evaporate.
Here's what Cursor assumes: you want to code. Replit? You want to ship. But Claude Code starts somewhere else entirely. It assumes you have a problem. Yes, the terminal looks technical because it is. But when you only need to explain problems, not understand solutions, everything shifts.
Cloud intelligence meets complete local access. Your machine, GitHub, databases, system internals. One conversation touching everything the terminal can reach. Intent becomes execution. No apps between you and what you want built.
Feels like 1985.
Remember DOS? Nobody wanted it. The thing looked hostile, demanded you memorize everything. Yet Lotus 1-2-3 lived there. WordPerfect too. Every breakthrough application, all of them, started in that ugly terminal before Windows made things pretty. The future always looks technical first.
And here we are again. Everyone's waiting for the friendly version while we're already building in terminals. Though not by typing commands. We're having conversations that execute across any system.
Conductor, Terragon, Magnet understand this. They're building interfaces beyond chat, visual layers that keep terminal power but hide the black screen. Eighteen months until everyone has what we're using now, wrapped in drag-and-drop.
MCP makes everything weirder. Model Context Protocol allows systems that never talked to each other to finish each other's sentences. Google Drive flowing into terminal commands. Local files meeting cloud databases. Makes you realize every app you use is just a keyhole view of what's actually possible.
But right now, in this terminal phase, we're learning how AI actually creates. No app filtering intent. No product manager's assumptions. Raw conversation with intelligence.
The real bottleneck isn't technical knowledge anymore. It's imagination. Last week I spent twenty minutes describing an animation. Not because Claude couldn't build it. Because I couldn't articulate what I saw in my head.
DOS lasted three years. Those who mastered it built Microsoft, Oracle, Adobe. Not from memorizing commands, but from understanding the substrate everything would be built on. This terminal phase will be even shorter.
Most people see a black screen and think "coding." I see the early glimpses of an “AI” computer.
I’ve been thinking along the same lines: if the model is the OS, then the “first-party apps” (the Office suite equivalents) are things like Deep Research, Coding, computer and browser control. But this will expand fast into data analysis, navigation inside specific UIs like ERP or PM software, and external communication—agentic search over email and chat. There’ll be obvious boundaries (OpenAI probably won’t chase the AI gaming space), but also cannibalization, the way Netscape never saw Windows coming for their lunch. The future isn’t written yet, but you can already make out the contours.
I like your writing but as I use LLMs myself. I am starting to see the LLM edits too. 😅