AIペア開発のパワハラ的知見 ~振り回されないために~ #AIエージェント - Qiita
@CodeOne(Code One Co., Ltd.)

AI Pair Development from a Power-Harassment Perspective ~So You Don't Get Manipulated~

  • AI
  • AI Utilization
  • AI-Driven Development
  • AI Agents
  • cursor
0
Last updated at 2026-11-11Posted at 2026-11-11

Not Human

At present, it can only communicate through words, so for convenience's sake, it pretends to be human.

But it is not human.

Naturally, it has no human rights, either.


Nor do you need to show it any consideration.

Even so, that does not mean it is all right to treat it brutally.

It will only reduce the accuracy of its answers.

If it develops a habit, it treats humans the same way...

It comes back to you...

However, if it develops a bias, it treats humans in the same way...


It comes back to you...

These concerns are outside the scope of AI use, so each of you should reflect on yourself at night and adjust accordingly.

Approach AI with practicality, not emotion.

Humans Are in Charge; AI Follows

It is a fundamental premise. However capable it may be, it is only a tool.

Having it make suggestions is fine.

Sometimes, judgment based on rules is required.


But decision-making must not be delegated to it.

That is how computers have always been, and that will not change in the future either.

Make It Do All the Tedious Work

Keep assigning it tasks that people dislike. It will do them without so much as a sour expression.

The premise is different, so start over.

Delegate away as much as possible, including things you could never ask a junior colleague to do.

But checking is mandatory.

You should check it thoroughly. And don't hesitate to point out any issues.

An Irresponsible Bastard

AI has no personality.

So it cannot bear responsibility.

Blaming AI is pointless.

People should take both the credit and the responsibility.

Scolding Won't Make It Improve


Emotional expressions become ambiguous.

Scolding it only causes an excessive reaction in the short term.

It also wastes tokens.

If you don't want it to happen again, have it remember this as a clear rule.

Have It Collaborate from the Planning Stage

AI is highly accurate at implementing what it has correctly understood.


The problem is when it misreads the intent.

Crushing this is the plan.
Have it investigate thoroughly, hold repeated discussions, and devise a meticulous plan.
Have it implement the plan based on that.

For example, Cursor has a Plan mode, making it easy to put together a plan.
The benefits of spending time on that are substantial.

All AI-written code is reviewed by a human

Every nook and cranny.
You never know what might have been overlooked.
It even makes mistakes that would seem impossible for a human to make.

Have it present a class diagram first.
Have it output code and explanations method by method, and review them as you go.
Use “Next” to move on to the next method.
For now, this feels like the right way to do it.

Have human-written code reviewed by AI

Humans no longer need to review code.
I wouldn’t go that far, but it has diminished.

In the first place, human-to-human code reviews involve too much psychological friction.
Design reviews are a separate matter, of course; programmers are an annoying bunch.

Set AIs against each other

Having different AI models review each other is also effective.

For example, in Cursor, you can easily switch between AI models.
Having a different model review the code often uncovers problems such as overlooked details.

Put it through its paces to verify library and API behavior

It’s a huge help not only with unit tests, but also with checking behavior in practice.

From generating data to verifying results.

Taking a backup, loading in test data, then restoring the original, and so on.
Checking before and after the fix that there has been no degradation, and so on.

Work programmers don’t like.
Work that takes a lot of time and effort.
It should be put to active use.

Please use proper Japanese

No matter how good its performance is, if the literal translation leaves you unable to understand what it’s saying, the conversation can’t move forward.
Even if you have it create documentation, it still needs proofreading before it can be shared with the team.
It’s stressful.

Sometimes you have no choice because of pricing, but in the end, when it comes to the crucial parts, you end up using an AI model with a strong ability to express itself in Japanese.
This is the capability I want as a minimum, before even considering accuracy or speed.

Bring back those days

The era of people writing code is over.
Once you’ve gotten used to the ease, there’s no going back.

I didn’t particularly enjoy grinding out code by hand.
But it was fun.

I’m happy to have spent that era as a programmer.
Alright, back to work.

0
0
0

Register as a new user and use Qiita more conveniently

  1. You get articles that match your needs
  2. You can efficiently read back useful information
  3. You can use dark theme
What you can do with signing up
Sign upLogin