How I Lost Half a Day: Even in the AI Era, Humans Still Make Mistakes

A Small Regret

Today I am reflecting on a mistake.

I wasted almost the entire morning.

It was not a system outage. I did not delete any data. But for an engineer, realizing that half a day of investigation was based on the wrong conclusion is still painful.

The topic was an internal investigation into a system whose behavior changed depending on the permissions of a particular file.

I investigated it with Claude Code.

At first glance, it looked simple. In reality, several small constraints were tangled together, and a slight misunderstanding of one condition could change the conclusion itself.

Looking back, I should have noticed it at that point.

This was the kind of task where I should have used the most accurate model available.

I Started in My Usual Mode

Recently, I have often been using Claude Sonnet for my daily work.

It is more than capable enough for writing, structuring ideas, and doing light research.

So today, without thinking too much, I started the investigation the same way.

That was the mistake.

I actually already had a personal rule:

Use the strongest model for investigation work.

For ambiguous requirement analysis, deep technical research, or work that requires checking multiple assumptions, I should use the highest-performing model even if it costs more.

That rule exists because I have failed enough times to learn it.

But today, I did not remember the rule itself.

It was not that I broke the rule consciously. I simply forgot that it existed.

A very human mistake.

Familiarity Is Useful, but Dangerous

When I think about it, the root cause was not simply the difference in model performance.

It was something much simpler.

Familiarity.

When you use the same tool every day, you start skipping decisions.

I should have paused and asked:

• What does this task really require?
• How much accuracy is needed?
• Which model should I use?

Instead, I started with:

The usual one should be fine.

And once the human brain sets an assumption, it can take a surprisingly long time to notice that the assumption is wrong.

That is exactly what happened this time.

I continued the investigation while missing a subtle interpretation issue around the permission settings. By the time I noticed, the morning was gone.

Choosing the Right AI Is Also an Engineering Skill

I do not want this to be misunderstood.

This is not a story about Sonnet being bad.

It is also not a story about Opus 4.8 always being right.

Honestly, I still cannot say for sure that I would have reached the correct conclusion if I had used Opus 4.8 from the beginning.

Any AI model can be wrong.

That is the baseline assumption.

The lesson is elsewhere.

Even though this was an investigation involving small, precise conditions like file permissions, I abandoned the effort to maximize accuracy from the start.

Nobody brings a kitchen knife to surgery.

It may still be able to cut.

But a reasonable person chooses a scalpel from the beginning.

This time, I did not.

Humans Are Still Responsible in the AI Era

Modern AI is genuinely powerful.

That is exactly why it makes us careless.

Situations that once would have made me think, “I need to be careful here,” can now quietly turn into, “AI will probably figure it out.”

But the final responsibility does not belong to AI.

It belongs to the human.

We cannot prevent models from making mistakes.

But we can control whether we tried to raise the probability of success from the beginning.

My reflection this time is not:

I should have used Opus and everything would have been solved.

It is simply this:

This was an accuracy-sensitive investigation, and I did not take the action that would have increased the odds of success.

An Expensive Lesson

Half a day of work disappeared.

Honestly, it did not feel good.

Still, if I do not repeat the same mistake, I can treat it as tuition.

From now on, before starting an investigation task, I want to ask myself just once:

Does this work really require high accuracy?

Is the model I am using appropriate for this task?

In the end, even in the AI era, what matters is not only the tool itself.

It is the judgment of when, where, and how to use that tool.

Today, I spent half a day remembering that.

I would have preferred a cheaper way to learn it.

And for the record, I used Opus 4.8 to write this article. That does not guarantee that this reflection is correct.