Claude Code Rolled Back My Productivity: Is Japanese Subject Omission Risky in the AI Era?

Today, Claude Code got the better of me while I was developing something.

To be fair, Claude Code was not really at fault.

I probably was.

I had received review comments from another department. I copied the comments exactly as they were, pasted them into Claude Code, and asked, “Do you understand?”

In my head, that meant something like this:

“These are review comments from another department. First, understand what they are saying. Then assess whether they are valid, and if necessary, let us consider how to respond.”

But I had not written any of that.

From Claude Code’s perspective, I had suddenly pasted a list of requested changes and asked whether it understood them.

Of course it understood.

And it did not stop at understanding.

It began changing the code at extraordinary speed.

By the time I realized what was happening and said, “Wait, wait. I have not asked you to fix anything yet,” it had already changed a considerable amount of code.

Fortunately, nothing had been committed, so I reverted all the changes and avoided any lasting damage.

As I watched a large set of changes disappear from the screen, a thought occurred to me.

Perhaps this was not just an AI problem. Perhaps it was also a Japanese-language problem.

In Japanese, Conversation Works Perfectly Well Without a Subject

Japanese speakers omit subjects all the time.

“Did you go yesterday?”

“Yes.”

“How was it?”

“Pretty good.”

This is a perfectly normal conversation. A great deal of information, including who went where and for what purpose, is missing from the actual words. Between people, however, the context fills in the gaps.

In fact, it would sound unnatural to spell everything out every time:

“Did you go to that restaurant yesterday?”

“Yes. I went to that restaurant yesterday.”

In Japanese, omitting what everyone already understands is often the more natural choice.

I had carried that same instinct into my instructions to AI.

In my mind, all the relevant premises were present:

• These comments came from another department.
• We had not yet decided to accept them.
• The first step was to evaluate their content.

But none of those premises appeared in the prompt.

A person might infer them from the flow of a meeting or the circumstances surrounding the conversation. Claude Code had no access to any of that.

As a general rule, information that is not written in the prompt does not exist for the AI.

More Capable AI Can Be More Dangerous

Older AI systems would often respond to vague instructions by asking, “What exactly would you like me to do?”

That could be frustrating, but it was also safe.

Today’s coding agents are different. They can absorb a substantial amount of context, infer what the user wants, locate relevant files, modify code, and even run tests on their own.

That autonomy is precisely why we use them.

But the more independently an AI can operate, the more damage it can cause when its initial interpretation is wrong.

If I were to compare AI to an employee, an older AI was like a new hire who waited for detailed instructions. A modern coding agent is more like an extremely capable employee who listens to half the explanation, says, “Understood,” and immediately starts working at full speed.

It is highly capable.

It works quickly.

But if it starts running in the wrong direction, it does so at a frightening pace.

That is exactly what happened to me.

English Exposes My Own Ambiguity

I use both Japanese and English regularly when giving instructions to AI. In my experience, English sentence structure is particularly effective when the goal is to communicate information unambiguously and make the logic explicit.

I notice this when writing blog posts as well.

When I write in Japanese, I omit subjects quite frequently. But when I ask AI to translate the text into English, the sentence often cannot work until a subject has been chosen.

Is it “I”?

Is it “we”?

Is it “they”?

Is it “the company”?

Is it “the government”?

Where Japanese can sometimes pass through an ambiguous point without resolving it, English is more likely to demand an answer to the question, “Who is doing this?”

When I read the resulting English, I sometimes think, “Yes, that is what I was trying to say.”

English often requires the relationship between the actor and the action to be made explicit in the structure of the sentence. Japanese, by contrast, can remain completely natural even when the subject is absent.

Between people, that is usually fine.

The problem arises when the other party is an AI.

When giving instructions to AI, it is not enough to say, “Fix this.” The prompt should also explain who raised the issue, what the AI is expected to do, and what it is allowed to change.

Although prompts are written in natural language, they share certain characteristics with programs. A program needs variables, targets, operations, and conditions. If they are missing, it cannot behave as intended.

As AI has become more capable, it has also become better at filling in those gaps and taking action anyway.

In this case, that convenience became the cause of the accident.

Language Reflects the Distinctions a Culture Has Chosen to Make

This leads to a broader and fascinating point about language.

Different cultures make fine distinctions in different areas.

Japan, for example, has a long history of eating fish. Its vocabulary distinguishes not only species, body parts, and preparation methods, but in some cases even the stages of a fish’s growth.

Yellowtail can be called buri, hamachi, or inada depending on its stage of development, and the terms vary further by region.

Fish mattered enough in Japanese life to require that level of distinction.

English, meanwhile, has highly specific vocabulary for livestock.

A sheep may be a ram, a ewe, or a lamb. Cattle may be described as a bull, cow, calf, or steer, depending on sex, age, and whether the animal has been castrated.

Just as Japanese speakers developed fine distinctions for fish, societies where pastoral farming was central to daily life needed precise ways to distinguish livestock.

A society’s vocabulary reveals what it has historically considered important enough to tell apart.

Language is not merely a tool for transmitting information. It also preserves where its speakers have drawn boundaries in the world and what they have learned to recognize as different.

The ease with which Japanese conversation can proceed without explicit subjects is part of the communication style the language has developed over time.

Now, however, the long-standing convenience of “you know what I mean” occasionally causes accidents when the conversation partner is AI.

AI Can Read Between the Lines, Until It Cannot

Modern AI is remarkably good at inference.

Even when a sentence lacks a subject, it can often reconstruct the intended meaning from the preceding conversation with surprising accuracy.

That is why the omission usually causes no trouble.

And because it usually causes no trouble, we gradually become less precise.

“Take a look at this.”

“Fix it.”

“The thing from earlier.”

“What do you think of this?”

AI can act on instructions like these. After they work one hundred times, we begin to feel that the AI understands us.

Then, on the hundred-and-first attempt, it says “Understood” while interpreting the instruction in an entirely different way.

And today’s AI can execute that misunderstanding at extraordinary speed.

In writing intended for people, omission and implication are part of what makes Japanese expressive and enjoyable. A blog post would be exhausting to read if every sentence said, “I went to a Zen temple yesterday. I walked through the temple grounds. I thought about Zen.”

Instructions for an AI agent are different.

There are several points that are better stated explicitly:

• Who provided the information?
• Is it an instruction or reference material?
• What should the AI evaluate?
• How much is the AI allowed to change?

When assigning work to AI, it helps to write Japanese a little more like writing a program.

Today, all I needed to write was this:

“These are review comments from another department. Do not implement them yet. First, evaluate whether they are valid.”

I omitted that one clarification. As a result, Claude Code performed a large amount of unnecessary work, and I then performed the additional work of deleting all of it.

We live in an era when AI can dramatically increase productivity. Yet one omitted subject or missing premise can send that productivity into negative territory.

At least today, the thing that needed an upgrade before the AI did was my prompt.