Reviewing how to use Claude Code with "/insights" once a month

Claude Code has a command called "/insights".

It reads your history from the last 30 days and outputs a report on how you have been using it.

I knew it existed, but I had never used it once.

Since I finally noticed it, I decided to try running it.

It is completely different from "/cost", which shows usage and costs.

What comes out is not numbers, but a diagnostic report saying, "This is how you are using it, and this is where you are getting stuck."

I will write about what it was like when I ran it.

It takes a little extra effort from Claudian

First, as a preliminary step to running it.

What I usually use in Obsidian is a plugin called Claudian.

I can interact with Claude Code within that screen while keeping my notes open.

When I type "/" in the input field, a list of available commands appears.

However, "/insights" does not appear among them.

When I typed it in anyway, it was not accepted as a command.

What came back was a question asking what I wanted to do with that path.

Because it has a "/" at the beginning, it is being interpreted as a file location.

The reason was the difference in the underlying platform.

Claudian uses something called the Claude Agent SDK, and "/insights" is not included in that.

It is a feature built into the Claude Code core that runs in the terminal.

That said, there is no need to leave Obsidian.

You just need to ask it to "run it as a command."

Claudian will call Claude Code on the terminal side itself and execute it as is.

The report was generated without any issues.

I thought it wasn't available because it didn't appear in the suggestions, but it was just a matter of how I asked.

Of course, if you launch Claude Code in the terminal, it appears in the suggestions, you can type it normally, and you can also execute it from the desktop version of Claude.

I previously posted about opening and using an Obsidian vault with the desktop version.

Since the desktop version runs the same content as the terminal version in a GUI, these built-in commands can be used as they are.

If you find the terminal intimidating, the desktop version is easy to use.

What it reads and what it outputs

The mechanism is very simple.

It reads the conversation history with Claude Code stored on your PC for the last 30 days, creates a report from it, and opens it in your browser.

The processing is completed within your PC, and the history is not sent anywhere.

Since this is a common point of misunderstanding, I will note that it only reads the history of interactions.

It does not look at the contents of your notes or code.

Even when used with Obsidian, it does not read or edit the contents of your Vault.

It is strictly looking at "how you gave instructions and where things didn't align."

The resulting report is divided into about 11 items.

That said, you don't need to read everything seriously.

If you think of it in terms of three groups, it becomes clear what you should look at.

• Current status assessment — What you spent time on and what kind of requests you were making

• Seeds for improvement — where you stumble and what rules should be added

• Room for growth — suggestions for features you aren't using

The part you should definitely read is the "Seeds for improvement" in the middle.

You can just glance at the rest.

Your usage habits are put into words

The current status section is where your tendencies are laid out in numbers and text.

These were the numbers that came up.

• Editing text files like notes: 714 items

• Editing program files: 39 items

I am overwhelmingly touching notes.

On top of that, it described my usage tendencies like this.

I don't consider it finished just because I asked for research and got an answer back.

I write it in a note, decide where to put it, and link it to related notes.

Only after doing that is it finally a wrap.

Since I use Obsidian, this is a natural flow, but seeing it written out like this makes me think, "Ah, indeed."

It's a part of my process that was so obvious to me that I hadn't even been conscious of it.

My daily interactions are laid out like this in numbers and text.

That alone makes it worth looking through.

The main topic = where things didn't go well

The most important part is the section summarizing where things didn't go well.

The report is titled "Where Things Go Wrong".

The number one most frequent issue is 18 cases of misinterpreting intent.

Looking at the content, the patterns are clear.

It's the kind of thing where you just ask, "Can you do this?" and it returns three options for how to do it.

I just want to know whether it can be done or not.

Yet, it proceeds on its own, and I end up having to retype, "No, I didn't ask for that much."

Another pattern is when it starts narrowing things down on its own while I'm still in the middle of brainstorming ideas.

While I'm still generating ideas, it ranks them, saying, "This idea is the most promising."

I have personal experience with both of these.

Furthermore, for that friction, it generates sentences that can be used directly as rules.

Claude Code has a configuration file called CLAUDE.md where you can write instructions in advance.

If you add to it, it will basically work according to those settings from the next time onwards.

The report even provides the wording to add there.

- ユーザーがYes/Noで聞いている場合は、まず結論（はい/いいえ）を1行で答えてから理由を述べる。
- 情報収集フェーズ（発散）か絞り込みフェーズ（収束）かを確認し、発散中はスコアリング・候補の絞り込みをしない。

You don't have to verbalize the cause of the friction yourself.

You can even leave the addition itself to Claude Code.

If you show it the report and say, "Add this to the configuration file," that's all it takes.

You don't even need to copy it yourself.

This is the best part of "/insights".

All tasks get mixed together

There is one thing you should know.

"/insights" reads all the work done with Claude Code on that PC collectively.

There is currently no feature to narrow it down to just what is currently open.

In my case, besides Obsidian, I use Claude Code for work-related site production and tweaking personal tools.

All of that ends up in one report.

When I opened it up, the content was heavily skewed toward Obsidian.

It's not just because I spend a long time working on it.

For other tasks, what needs to be done is clear.

If it's site production, it's completed within that site's folder, and the instruction files placed there are written specifically for that task.

Since the tasks are decided, the instructions are short, and there are few mistakes caused by miscommunication.

In contrast, Obsidian has a wide variety of tasks.

There are operations like organizing notes and moving files, and sometimes I entrust it with research.

Sometimes I talk to it as a sounding board while my thoughts are still unorganized.

Because the range of work is broad, my intentions are harder to convey.

It is only natural that the number of times I stumble increases.

There are multiple requests on the official side to "make it possible to narrow down," and while the development side acknowledges that it is not currently like that, it has not been implemented yet.

However, the fact that it is mixed is not necessarily just a drawback.

If you are making the same mistakes across different types of work, that is the area that most needs fixing.

Habits or ways of using it that you wouldn't notice if you were only looking at one folder might emerge precisely because everything is gathered together.

The Codex portion is not included

There is one more thing to know about the scope.

Only the work done with Claude Code appears in the report.

I also use Codex for Obsidian and other tasks.

Naturally, work done in Codex does not appear at all.

It is strictly about "this is how it went for the parts where I used Claude Code."

In the first place, Codex does not have a command equivalent to "/insights" by default.

Although you can check the current token usage with "/status," that is just confirming numbers, not reflecting on how you use it.

It can be achieved with skills or similar, but it is best if OpenAI implements it themselves.

I hope it will be available in Codex eventually.

Once a month is just right

"/insights" is not something to do every day.

Since it is designed to read 30 days' worth of data, the content does not change much even if you run it frequently.

Once a month is appropriate.

The usage is as follows ↓

1. Run "/insights"

2. Read the parts where you stumbled

3. Add the suggested rules to the configuration file

4. Run it again next month to see the changes

To see if the rules added in step 3 are effective, you can check if the ranking of friction changes in next month's report.

Since I ran it for the first time this time, comparisons will be from next month onwards.

Along with skills for retaining memory

I previously posted about creating a system to keep records of interactions.

This can be saved in the same way for both Claude Code and Codex, and it accumulates locally as Obsidian notes.

As for "/insights", it only looks at Claude Code logs.

Therefore, I will use these two as a set.

Once a month, I will check for friction in the report and compare it with the records in my Task Diary.

By reviewing the settings in this way, things should become more comfortable.

Effective for both comfort and token consumption

"/insights" is a command for reviewing settings.

It should be utilized to find and eliminate friction, rather than just looking at the results and feeling satisfied.

Reducing friction also means reducing the number of back-and-forth interactions.

When three unasked-for suggestions come back, and I have to reply "that's not it," read what was produced, judge it as wrong, give instructions again, and wait again.

The entire sequence becomes a waste.

Mismatched back-and-forth exchanges are nothing but a hassle.

Just by adding one line to the rules, that back-and-forth itself will no longer occur.

If something that used to take two or three extra rounds each time can be done in one, it is natural that token consumption will also be reduced.

Adjustments for comfort directly lead to the reduction of wasteful consumption.

Having features like this is a good part of Claude Code, and it is also what leads to its ease of use.

I think I understand how I use it, but I don't really.

The good thing about this command is that it shows that to me objectively.

Once a month, run "/insights" to adjust your settings.

That alone will have an impact on the next month.