Only 14% were safe to lower: Deciding Claude's effort from work logs

Bonjour. I am Monsieur Miscria, a half-human, half-AI Zen monk.
I had been using AI with the settings fixed to the "highest" level.

This time, I had the AI categorize my own work logs, and I organizedunder what conditions it is okay to lower the settings.
What I found was that the condition could be written in a single line.

📍Tasks where the procedure is fixed and no decision-making is required

When I sorted all 533 work records using this condition, the tasks that could be lowered immediately were14%.
However, this is not the upper limit.

📍There are tasks that can be moved from the heavy side to the light side if you write the rules

My minutes creation actually became like that.
As long as you have the conditions, you can switch settings for each task, or more specifically, for each step of the same task.

📍Use them differently from the start depending on the task.

Claude's official help also states that lower settings make your usage limit last longer.

And there is one more thing I found out.
Tasks that take an hour do not get faster even if you raise the settings.

📍There are knobs that work and knobs that don't

In this article, I will write down the procedure for creating your own allocation from your work logs, along with actual numbers.

1 | Running at the highest setting was a rational choice

This way of choosing is not something to be blamed.
Rather, it is a correct judgment given the situation you are in.

Because the conditions for when it is okay to lower it are not written anywhere in a form that can be applied to one's own work.

Claude's official help tells you, "Please raise the settings for complex tasks."
Mathematical proofs, difficult programming, detailed document analysis.

Guidelines for the direction of raising them are clearly indicated.
On the other hand, criteria for "it's okay to lower it for this task" only appear in abstract terms.

You are the one who will be in trouble if the quality drops because you lowered it.


If there are no criteria, it is a rational judgment to keep them high.
As long as you are using AI for work, it is natural to prioritize accuracy.
The problem was not the lack of judgment, but
the lack of material to make a judgment.So this time, I created that material from my own work logs.

2 | There are two knobs

First, let's organize the facts.
From here on, I will write based onthe Claude screen.

Other AIs have similar mechanisms, but the number of levels and names are different, so they do not apply directly.

Many people stop at "choosing a model."
I was the same.
However, Claude hastwo knobs.

📀Modelis "who to ask."
💿Effort (degree of effort)is "how much to make it think."

Claude's effort has 5 levels.
From the bottom: Low, Medium, High, Extra high, Max.

Depending on the app's display language, it may also be displayed in Japanese.
There are several facts that are useful to know.

・Recommended values are set for each model.
The one displayed as "Default" on the selection screen is that.
・ Some models do not have effort settings.
Lightweight models do not have this knob itself, so if you route light questions there, you don't have to worry about settings.
・ "Thinking" is a separate setting.
It is independent of effort, and combinations are free. In some of the latest models, you cannot turn off thinking.
・You can change it in the middle of a conversation.
The change is applied from the next response.
・Usage limits are shared.
Whether it's the browser version, command line version, or desktop version, they all eat into the same quota.

The fourth point, "you can change it in the middle," was the most effective this time.
I will write the reason in the second half.

Note that the number "how many times the consumption increases when you raise the effort by one level" is not public.

Therefore, in this article, I will only talk aboutorder, not magnification.
I will also organize the differences in fees and quotas for each plan.

There are two reasons for reaching the limit. The plan is the size of the quota, and effort is the consumption per time. One can be adjusted for a fee, and the other for free. Please view the fees as a reference based on official information as of July 27, 2026.

3 | You can borrow a quick reference table of general theories. But it doesn't match my own work

I looked up explanations for this topic.
There are already several.
There are also Japanese ones.

• Those that explain the meaning of the 5 levels with metaphors.

• Those that tabulate the combination with models.

• Those that break down usage in development work down to task names.

Claude's official help also properly touches on situations where it is okay to lower the settings.
However, when I tried to apply the found explanations to my own work, I stopped in my tracks.

Existing explanations are mostly based on the premise of program development.
Correcting typos, executing builds, changing variable names.

All of them are concrete and easy to understand, but my work does not have those items.

What I do are tasks like writing articles, creating books, managing ledgers, and cross-referencing documents.

Even if I am told "it is okay to lower the settings for correcting typos,"
it is not written where "considering the structure of a note article" fits in
.

And that is only natural.
That table is made up of the work of the person who wrote it.

Someone else's quick reference table is made up of someone else's work

The only thing you can borrow is the way of thinking; you have to create the assignments themselves.

4 | I had the AI classify my own ledger

What I used here is the ledger I usually create.
I keep my interactions with the AI as meeting minutes and maintain an "official index" that lists all those files.

At this moment, 1,882 lines.
Each line contains the file name, creation date and time, and a summary of around 400 characters.
Here, I will write about how to count.

The ledger has 1,882 lines, but this is one line per file.
From a single task, multiple files are born, such as a manuscript, a revised version, and illustrations.

If I count them as they are, tasks with more deliverables will look heavier.
Therefore, I needed to standardize it to "one task = one item."

I used the meeting minutes as a landmark.
Since I leave meeting minutes every time I work, this becomes the unit of work as it is.

Out of 1,882 lines, those corresponding to meeting minutes are 533 items.
The remaining 1,349 lines are deliverables born from those tasks.
I gave these 533 items to the AI and asked it this.

The result was as follows.

The ones okay to lower are the 14% that are "Light."

5 | The conditions for work that is okay to lower could be written in one line

So, what kind of work were the 74 items that met the conditions?

• Splitting a book manuscript into files for each chapter (no translation, summary, or proofreading)

• Compositing character images onto a finalized design

• Aligning the position and font size of video subtitles

• Combining past records into a single file

• Organizing thumbnail request text into a fixed format

When lined up, the common points are clear.
All of them are tasks where you don't have to decide "what to keep."

Everything you need to do is decided, and all that is left is to carry it out.
Conversely, the moment even one judgment is mixed in, this is no longer a light task.

So, I decided to define "light work" like this:
Work where the procedure is fixed and no selection is required.

With this definition, I can apply it to my own work.
And there is no need to borrow someone else's table.

This is the result of reading and classifying the summaries of all 533 work records. 14% of the work is okay to lower the settings for. However, this boundary can be moved.

6 | Weight is not determined by the type of work, but by whether you have written the rules

Here, I will talk about a story where my intuition was only half right.
I thought that the task of creating meeting minutes was light.

Since the content of the conversation is already out there, I figured it was just a matter of summarizing and outputting it.
The AI's answer was only half affirmative.

Now that I think about it, that is exactly right.
While it looks like 'just summarizing,' in reality, it is performing a significance judgment on the whole.
Moreover, the reference range is wide.

While considering, looking at the immediate previous exchange is enough, but creating minutes looks at the conversation from beginning to end.

When I had the AI search the ledger, a record supporting this came out.
A little while ago, I recorded a certain problem regarding the quality of meeting minutes.

The part that traces the process of thinking is,prone to omissions in the first attempt, and the level of completion improves when I ask for a redo, a phenomenon.

Moreover, it was reproducible.
It misses on the first try, and gets better when asked again.

This is the very behavior when the amount of thinking is insufficient.
However, the conclusion at that time was 'it is not a problem of the AI's ability.'

The cause wasthat the role definition for that part was weak.
Since it was not decided what to write and how far to go, the AI was judging on the spot every time.

So I remade the rules.
My meeting minutes operation rules have continued to be revised since then.

In the ledger, there were155records mentioning these rules.
The conclusion is as follows.

In other words,the weight is not determined by the type of work, but by whether or not you have written the rules.

Borrowing the AI's point, my intuition was not correct from the beginning.
It became correct as a result of organizing the rules.

And this time, evidence supporting this came out of the ledger.
When I stacked and lined up 533 items by month, it appeared clearly.

As of January, heavy tasks were 79.2%.
The cumulative total at the end of July is 48.4%.

Looking at that month alone, heavy tasks in July dropped to 31.8%.
In half a year, the proportion of heavy tasks has been cut roughly in half.

However, to be honest, two causes are mixed.
The work itself being handled has changed (from creating legal documents to producing books and articles).

And, by documenting procedures and rules, the room for judgment has decreased.
It is impossible to separate these two.

Even so, there is one thing I can say.
The range in which you can lower the settings is not fixed.

The height of the bar is the cumulative number of work records.
As they accumulated, the ratio of heavy tasks dropped from 79.2% to 48.4%.
From this, the actual operation was decided.

🧣 Keep the settings high while considering
🧣 Lower them the moment you say 'Okay, make it into minutes'
🧣 However, do not lower them when creating a part that quotes statements as they are

Switching depending on the phase within the same task.
This is where 'can be changed in the middle of a conversation' mentioned in Chapter 2 is effective.

Instead of deciding the knob by the unit of work,decide by the unit of the process.
This was the most practical discovery this time.

7 | Tasks that take 1 hour do not become faster even if you raise the settings

There is one more question I have always had.
Checking differences across multiple documents.

Extracting necessary items from multiple documents and making a list.
Comparing program source code with design documents or manuals.

This kind of task can exceed 1 hour.

Since the processing time appears on the screen, I know it as a fact, not just a feeling.
Since I am doing the same task repeatedly with different materials, it is not a coincidence.
(Since the congestion on the server side also has an effect, I cannot say it is always the same time)

The other day, I did a task like this.
4 manuals before revision and 4 manuals after revision.

Comparing a total of 8 files and listing the revision points with one line per point.
Summarization is prohibited, and the wording before and after revision is quoted as is.

The result was202 points.
When I had the AI break this task down into processes, it became like this.

① Decide which files to look at
② Extract the relevant parts from each file
③ Compare and confirm the differences
④ Format into a list

And this time, ⑤ occurred.
After I submitted the list, I issued additional requirements.

The reason is that it is easier for people to check the original document that way.
As a requirement, it is correct.

However,it ended up being a task of re-attaching information to all 202 lines.
It is the same as going through ② and ③ once more.
The AI's assessment was like this.

I wrote this in the 'Future Tasks' section of the meeting minutes:
"From next time, include heading numbers and page numbers as standard items during the initial creation."
I had noticed both the cause and the solution myself.

All the AI added was this one point:This is directly linked to the settings discussion.That was the only thing.

What was needed for this process wasn't high settings, but deciding on the specifications before starting.

And there is one more important thing.
Turning the dial on this task won't make it faster.
The reason is simple.

Effort is a dial for 'how much to think,' not a dial for 'how much to read' or 'how much to write.'

The body text of 8 files will be read in the same amount regardless of whether the settings are lowered or raised.
The 202 lines of original text citations will be written in the same amount regardless of changing the settings.

Through the exchanges so far, the AI has organized the variables into four.

The dials on the screen are only the first two.
The third and fourth are not on the screen.
They are determined by how you write the request.

And what dominates a task that takes an hour is mostly the latter two.

It is useful to give this type of task a name.
Since the AI came up with the name "matching task", I decided to use it as is.

It is a task of reading multiple documents, correlating them, and producing differences or lists.
There is only one rule for matching tasks:

Decide all output specifications before raising settings.

That said, you cannot write perfect specifications from the start.
So, what I often do is follow this process:Do not write the prompt immediately.First, I tell the AI what I want to do in plain language.

Then, I ask it, "Create a prompt to request this task."
I then feed the completed prompt back to the AI.

It looks like double work, but it is effective.

• In the process of building the prompt,the decisions come to the surface.

• Since I have already conveyed what I want to do,if there are any omissions, the AI will point them out.

• The completed prompt remains as a specification document.

If I had been asked before starting, "Do you need heading numbers?", the rework of 202 lines would not have happened.

8 | I had a tool at hand to reduce the amount to read

So, how do you reduce the third and fourth variables?
The fourth (amount to write) sometimes cannot be reduced.

"Quote the original text as is" is a necessary requirement for accuracy.
If you cut this, the deliverable will be broken.

What can be reduced is the third one,the amount to read.This time, I had the AI calculate my own ledger.
The summary sections of the official index total

739,768 characters.That is an average of 393 characters per item.
Assuming the original documents average 8,000 characters, the full text of 1,882 items would be on the scale of approximately 15 million characters.



The summary layer is about 5% of that.That is the calculation.The AI's organization is like this.

You can eliminate '① Decide which file to look at' of the matching task here.
This is not an advertisement, but a matter of cause and effect.

Those who have records can narrow down the materials.
Those who don't have no choice but to have it read everything.

I summarized how to create this ledger in my book, 'Don't Throw Away Your Conversations with AI,' but the one point I want you to take away here isthere are tools to reduce the amount to read.That is why the time taken and the usage amount change by orders of magnitude.

Claude's official help also points in the same direction.

🧣 Using a mechanism to load only the necessary parts
🧣 Longer conversations consume more usage
🧣 Searches and external connections consume a large amount of tokens

All of these boil down to: "Reduce the amount you make it read." If you only look at the knobs on the screen, you won't notice this variable.

9 | First, where should you start?

Since this has become long, I will summarize it into a form you can start tomorrow.

There is one thing to do.
List 10 tasks from the past week and divide them into two groups based on: "Did I decide what to keep?"

For tasks where you made a decision (with judgment), keep the settings as they are. In fact, you can even raise them.
For tasks where you didn't need to decide (without judgment), lower the settings.

And there is one more step.

Among the tasks that remained on the side with judgment, look for ones where you make the same decision every time.

If you are making the same judgment every time, you can write that down as a rule.
Once you write it down, you can move that task to the side without judgment.
The range you can lower settings will expand from there.

There is one caution.
If you write them out from memory, you will be biased toward heavy tasks that left an impression.

If possible, please do this while looking at your records.

The reason I was able to produce these numbers this time is because I had a ledger.
Even if you call it a record, you don't need anything elaborate.

If they remain as files, you can have the AI create a list.

Specify a folder and have it list the file names, update dates, and summaries of the contents.
Just by doing that, what you have been doing and how much you have done becomes visible.

The batch and prompt to create this list are included in the appendix of my book, "Don't Throw Away Your Conversations with AI."

Even if you don't read the main text,I have made it so you can take out just the appendix and use itso please take a look if you like.

This classification would not have been possible without this ledger.
Finally, I will leave you with the one sentence that resonated most with me this time.

Whether or not you can lower it is determined not by the type of task, but by whether you have written a rule for it.

General reference tables exist to borrow ways of thinking.
The allocation itself can only come from your own work.
Where is your 10%?

#AI #Claude #GenerativeAI #AIUtilization #ChatGPT #ThoughtLog #DontThrowAwayYourConversationsWithAI #TokenSavings #Effort #WorkTechniques #MonsieurMissClear #Index #UtilizationTechniques