How to Train AI Agents to Work Smarter: Building the Right Rules for Better Results

Introduction

Hello, I’m Mia Sato, AI Researcher at GDX.

One question I hear more often these days is, “I tried asking an AI agent to do some work, but it did not behave the way I expected.”

I had the same experience at first.

I would ask it to do too many things at once, then look at the result and feel disappointed.
I repeated that pattern several times.

But after trying it a few times, I realized something important.

An AI agent is not just something you command.
It is something you train.

That does not mean it naturally becomes smarter just because you use it many times. If the way you teach it, or the rules you give it, are unclear, it may repeat the wrong action or the wrong judgment standard. Then you may spend even more time fixing the result later.

In this article, I will share the common mistakes people tend to make when they first start using AI agents, what can happen when the training goes wrong, and how to create an environment where AI can learn the right way to work.

At the end, I will also share two examples I actually tried in my own work.

Common mistakes when training an AI agent

First, I would like to share some mistakes I made myself while using an AI agent.

1. Asking it to handle too many tasks at once

In the beginning, I often gave instructions like, “Collect the data, summarize it, create a chart, and compare it with last week.”

But when I packed too many steps into one request, some steps were skipped. Sometimes an early mistake also spread into the rest of the workflow.

Since then, I have tried to split the work into smaller steps and check the output along the way.

2. Treating the AI’s guess as the correct answer

I also gave vague instructions such as “Enter the amount” or “Summarize this nicely,” and assumed the result was correct if it looked reasonable.

But even when the result is correct, it does not always mean the AI understood the right rule.

If the conditions change slightly, it may not make the same decision again. So now I try to check why the AI selected a certain item, not only whether the output looks right.

3. Letting it process many items before checking the first one

I once thought, “This is the same type of task, so it should be fine,” and asked the AI to process several items at once from the start.

But if the first item is handled incorrectly, the same mistake can be repeated across all the remaining data.

Now I ask it to process one item first. Once I confirm that the input method and judgment criteria are correct, I let it continue with the rest.

4. Giving slightly different instructions every time

As I adjusted my instructions for each task, I sometimes ended up giving slightly different rules every time.

When that happens, the AI may not know which standard to follow. Even under the same conditions, the result can change.

These days, I try to separate “basic rules,” “exception rules,” and “instructions for this time only.”

5. Mixing temporary instructions with normal rules

There were also times when I said things like “Do not enter it this time” or “Stop before sending this time” for testing, but did not clearly say that it was only for that test.

If temporary instructions and normal rules get mixed together, the AI may continue using the same behavior next time.

So for test instructions, I now make it clear by saying “This time only” or “Do not use this as a normal rule.”

Through these experiences, I realized that it is more important to reduce guesswork, align judgment criteria, and create an environment where mistakes can be corrected early, than to simply make an AI agent remember many tasks.

Rules that make it easier for AI to learn

So, how can we ask an AI agent to work more smoothly?

I think the key is to treat it as if you were training a new team member.

Instead of only explaining the steps, it helps to define the goal, the judgment criteria, and what the AI should do when it is not sure.

Once these rules are in place, the AI becomes less likely to get lost, and mistakes become easier to catch early.

Example 1: Filling in weekly work hours in CrowdLog

The first example is weekly work-hour entry.

This is the task of entering this week’s work hours by project in CrowdLog.

At first, it is tempting to simply say, “Please enter all of this week’s work hours.”

But with that instruction alone, the AI does not know which project should receive how many hours on which day. It may get stuck.

So I separated the goal from the details.

I first told it, “I want to fill in the work-hour sheet for this week, Monday through Friday, and bring it to the point right before submission.”

Then I gave the actual inputs, such as “Project A is three hours each on Monday, Wednesday, and Friday. Project B is four hours each on Tuesday and Thursday.”

For the first day, I checked the input together with the AI. Once I confirmed that the entry method was correct, I let it continue with the remaining days.

At that time, the screen also showed projects with similar names.

If I had not checked the first day, the AI might have continued entering the same hours into the wrong project.

So I added one more rule: “Only select a project when the project name matches exactly. If it does not match, do not guess. Stop and ask for confirmation.”

Once I shared not only the input method, but also the selection condition and stopping condition, the same type of task became much easier to delegate.

I still pressed the final submit button myself after checking the content.

The weekly CrowdLog screen where work hours are being entered. The orange arrows on the screen show where the AI agent is operating while learning how to input the information.

Example 2: Entering receipt information into Rakuraku Seisan

The second example is expense reimbursement.

I tried asking the AI agent to transfer information from a receipt into the input form in Rakuraku Seisan.

Here again, if I simply say, “Please enter this receipt into Rakuraku Seisan,” the AI may not know which information should go into which field.

A receipt contains many pieces of information, such as the store name, date, amount, and notes.

So I clearly explained the mapping.

For example, “receipt date → usage date field,” “store name → payee field,” “tax-included total amount → amount field,” and “notes on the receipt → description field.”

I also gave a rule for cases where the account category might be unclear, such as using “travel and transportation expenses” when appropriate.

Some receipt images are hard to read.

On the first receipt, there was a moment where the AI seemed unsure whether it should enter the subtotal or the tax-included total.

Instead of only giving the correct amount for that receipt, I added a rule: “In the amount field, enter the final payment amount including tax.”

If I had only corrected that one answer, the AI might have repeated the same confusion with another receipt format.

Once the reading became more stable, I asked it to process the remaining receipts. Then I reviewed all entries myself before registering them.

The point was the same in both examples.

Share the goal first.
Specify where each piece of information should go.
Check the first few steps together.
Keep the final action in human hands.

Just by changing the way I guided the AI, the range of tasks I could safely delegate became much wider.

Conclusion

What I felt through this experiment is that the important question is no longer only “Which AI tool should we use?”

To delegate work to an AI agent in a stable way, we do not need to make it memorize one correct answer after another.

What matters more is making the judgment criteria clear.

If we organize the basic rules, exception rules, one-time instructions, and stopping conditions, the AI becomes more likely to make decisions based on the same standard, even when the situation changes slightly.

In that sense, the real point may not be simply “training AI.”
It may be designing rules that make it easier for AI to judge correctly.

As AI tools continue to improve, the ability to decide what judgment criteria to give AI, and where humans should step in, will become more important than individual features alone.

A good starting point is to choose one small task and put your own hidden judgment criteria into words.

I believe that this steady process will become the foundation for more stable collaboration between people and AI.

References

• Official: Claude / Anthropic / Anthropic
• Official: Get started with Claude in Chrome / Claude Help Center / Claude Help Center
• Official: Using Claude in Chrome safely / Claude Help Center / Claude Help Center
• Official: Using Claude App Intents, Shortcuts, and Widgets on iOS / Claude Help Center / Claude Help Center
• Japanese commentary: Trying Claude in Chrome for browser automation with natural language / DevelopersIO / DevelopersIO
• Japanese commentary: Practical guide to Claude in Chrome and advanced use cases / Motohiko Sato / note / note
• Japanese commentary: Complete manual for AI agent prompt design / Visionary Japan / Visionary Japan

※ Part of this article was created with the support of AI and edited by the author. The content reflects the author’s personal views and does not represent the official views or statements of GDX Inc. The information is provided for reference purposes only. Please refer to official announcements and primary sources for the latest details.