AI Is Our Car. Do Engineers Still Need to Train Their Legs?

The other day, Claude Code stopped working while I was in the middle of my job.

Errors appeared, and it would not connect properly. When I checked the service status, there was indeed an outage.

I do not depend exclusively on Claude Code, so I could move blocked work over to Codex. Sonnet was also functioning relatively well on the Claude side, so the work itself did not stop completely.

Even so, as I stared at the error screen, an unpleasant thought occurred to me.

If Claude Code, Codex, and every other AI tool became unavailable, could I still fix my current project by myself?

The answer is probably yes.

But there is no chance I could do it at the speed I once did.

After Driving Everywhere, Walking Started to Feel Like a Chore

Since I began using AI seriously for development, I feel that my abilities have changed slightly.

I can look at an architecture and recognize that something is wrong with the design. I can examine a proposal from AI and decide whether to accept or reject it. I can also notice when the implementation has drifted away from the requirements.

On the other hand, writing code while trying to recall small details of syntax, following logs one line at a time, and eliminating possible causes through repeated searches have clearly become tedious.

It is not that I can no longer do them.

I no longer want to.

It is like using a car.

Tell someone who drives everywhere that they must suddenly walk six miles today. They are probably capable of walking that distance, but it will feel extraordinarily far.

Debugging tasks that were perfectly normal before AI are beginning to feel the same way.

Should Engineers Schedule Regular “No AI Days”?

The serious response to this concern would be:

“Relying too heavily on AI will weaken your abilities. Once a week, you should write code entirely by yourself without using AI.”

It would be strength training for engineers.

I am not going to do it.

Nobody who can drive to a client visits them on foot once a week merely to preserve the ability to walk six miles.

There is no need to practice long division when calculators exist, or write Java in a plain text editor when an IDE is available.

It is natural for abilities made unnecessary by convenient tools to decline somewhat.

If I have that much spare time, I would rather use AI to complete other work.

Would Claude and Codex Really Fail at the Same Time?

There is also a more practical point.

An outage affecting Claude Code does not cause every generative AI service in the world to disappear simultaneously.

If Claude is unavailable, I can move to Codex.

If only one Claude model is affected, I can switch to a model that still works.

If those options also fail, there are others.

Of course, anything can happen if we imagine a massive network outage, but that becomes a different kind of problem.

Sometimes Waiting for Recovery Is the Correct Decision

Even if every AI tool were temporarily unavailable, continuing the work manually would not necessarily be the best choice.

Suppose a bug that once required ten hours of investigation can now be found by AI in ten minutes.

If the AI service is expected to recover in three hours, waiting three hours and then fixing the issue in ten minutes would still be faster.

I can do other work while I wait.

There is something admirable about saying, “A real engineer can do everything manually even during an outage,” but work is not a test of endurance.

When a train stops, walking along the tracks to reach the office is not considered a business continuity plan.

But You Should Still Keep a Map

That is why I have no intention of training myself to code without AI at this point.

There is, however, one thing I believe I still need.

I should keep an architecture diagram of the system I am responsible for.

Which services connect to which? What does each repository contain? Where does the data enter, and where does it go? Where are the critical processes, and where should I look first during an incident?

When AI handles the daily work, even these questions can gradually become something answered with, “I will ask Claude.”

That is slightly dangerous.

If I forget the finer details of writing code, I can look them up. I can also wait for AI to recover and ask later.

But if I no longer understand what I am building, I can hardly call myself an engineer.

It is probably inevitable that driving everywhere will weaken my legs.

That does not mean I intend to start running every morning.

But if the car breaks down, I should at least have a map that tells me where I am.