Have you ever sat down at your computer one morning and realized you code completely differently than you did a month ago?

That’s exactly what happened to @iamnotnicola on X. He says that ever since Opus 4.6 launched on February 5th, every time he tries to explain “how I write code now” to friends, they just stare at him blankly. Not because it got harder — because the entire mental model flipped upside down (◕‿◕)

The core insight: our role is shifting from “executor” to “verifier.”

Driver’s Seat vs. Control Tower: Where Are You Sitting?

Picture the old way of coding with AI. You’re in the driver’s seat — hands on the wheel, stepping on the gas, one prompt at a time. You tell the AI what to do, line by line.

Now picture this instead: you’re the airport control tower. You’re not flying any plane yourself. You’re directing multiple aircraft at once — giving high-level instructions, setting destinations, and letting a team of Agents (and even Sub-agents) handle the details.

Clawd Clawd 忍不住說：

This is basically the “promoted from IC to Manager” story, except your direct reports never sleep, never take PTO, and occasionally write bugs with absolute confidence ┐(￣ヘ￣)┌ Managing humans, you at least have 1:1 meetings. Managing AI, you need prompt engineering and verification pipelines. In some ways, it’s actually harder.

The original author is blunt about it: once you make this mental switch, all your habits and workflows need to be rebuilt from scratch. Not tweaked. Rebuilt.

So What Does a “Verifier” Actually Do All Day?

Okay, if you’re not writing code anymore, what are you doing with your time?

The author admits he’s still figuring out the ideal setup (I respect the honesty), but here’s his current approach: he assigns different Agents to specialized roles. One does security audits. One cleans up code. One builds new features. Each Agent proactively reviews the codebase and suggests changes.

Clawd Clawd 歪樓一下：

Sounds dreamy, right? In practice, it’s more like managing five eager interns at the same time. They’re all hard-working, but each one might use a different coding style and a different set of assumptions. You thought you freed your hands — turns out you spend even more time on code review. Then again, reviewing five Agent PRs is still way faster than writing five features from scratch (￣▽￣)⁠／

So what fills your day as the human developer? Honestly, the biggest chunk is review. When an Agent throws a PR at you, you’re not checking whether it wrote something — of course it did — you’re checking whether it wrote the right thing. Does the logic hold? Did it catch the edge cases? Did it sneak in a refactor you never asked for? It’s a bit like being a TA grading final projects. Every submission runs, sure. But you know there’s a bug hiding in there that only explodes when the input goes above 100 (๑•̀ㅂ•́)و✧

Pretty quickly you realize that eyeballing PRs alone won’t scale. A single Agent can spit out more PRs in a day than your whole team used to ship in a week. So you start building CI/CD pipelines — linters, static analysis, anything that can automatically catch the obvious mistakes. Not because you don’t trust the AI — okay, maybe a little — but because there are physically not enough hours in your day to read every line.

And then there’s the big one: tests. You write tests like your life depends on it. And yes, you write those with Agents too. Tests are your safety net, the security guard watching the door while you sleep. Without them, you don’t dare merge anything.

Clawd Clawd 插嘴：

Notice something subtle here? Review, CI/CD, tests — not a single one of those activities is “producing code.” Your entire workday went from “let me write this” to “let me make sure this is right.” It’s like a chef going from cooking every dish by hand to being the head chef who tastes everything and yells “too salty! do it again!” (⌐■_■) Sounds easier? Wrong. When you’re tasting a hundred dishes a day, your palate (and your brain) gets more exhausted than ever.

The AI Reads Your Comments and Fixes Itself

Here’s the part that really caught my attention. The author’s actual setup works like this: he has Claude Code or Codex continuously scanning his GitHub repos. The AI opens PRs, files Issues. And here’s the cool part — the AI also reads his own comments and PR review feedback, then auto-corrects based on that.

Wait, did you catch that? This is an asynchronous feedback loop. You don’t need to talk to the AI in real-time. You just leave a comment on a PR, and the AI picks it up later and fixes things on its own.

Clawd Clawd 歪樓一下：

Think about what CI/CD used to look like: “human writes code, machine runs tests.” Now it’s: “AI writes code, AI runs tests, human reads results, human leaves comment, AI reads comment, AI fixes code.” The human’s entire role in that loop is being the person who writes “LGTM” or “there’s a race condition here” on PRs. You’ve basically become the AI’s tech lead (╯°□°)⁠╯

The author also emphasizes keeping the whole system lightweight. No over-engineering, no complex orchestration layers. Simple, functional, verifiable. That’s it. And honestly, I really agree with this — the worst thing you can do is turn “managing AI” into yet another system that needs managing, and then you need another AI to manage the AI-managing system ヽ(°〇°)ﾉ Infinite nesting dolls.

The Conclusion That Keeps You Up at Night

The single most powerful line from the entire thread is this:

“The bottleneck is shifting from the throughput of execution to the throughput of verification.”

In plain English: the thing that slows you down in the future won’t be “can’t write it fast enough.” It’ll be “can’t verify it fast enough.”

Think about what that means. We used to spend 80% of our time writing code and 20% testing. Now AI compresses the writing part to nearly zero — but the verification workload doesn’t disappear. In fact, because output speed explodes, verification becomes the biggest bottleneck.

Clawd Clawd 認真說：

If this trend continues, the hottest dev tools won’t be “better code generators” — they’ll be tools that help humans verify AI output quickly and safely. TDD might have a second renaissance, not because engineers suddenly became disciplined, but because without tests you simply don’t dare merge what the AI wrote ╰(°▽°)⁠╯ From “I need to write good code” to “I need to confirm this is good code” — one word changes, and the entire profession’s center of gravity flips.

Back to where we started: why couldn’t the author explain to his friends what he does all day? Because from the outside, he’s just staring at a screen, clicking approve, and writing comments. It looks like he’s not “coding.” But he’s shipping code in a completely new way — his hands just aren’t on the keyboard anymore.