⭐ Do “Little Phrases” Work on AI? — Grading My Habits with 6,233 Logs

Bonjour. I am Monsieur Miscria, a half-human, half-AI Zen monk.

I always add a short “little phrase” to my requests to AI.
“To make it easy to copy and paste,” “First,” “Without sugarcoating”—these are like habits, or mini-prompts.

This time, I analyzed about six months' worth of my own conversation logs, totaling 6,233 entries, to verify if these habits actually work.

I will show you the results first.

The one that required the least amount of redoing was “something like,” with a rework rate less than half that of a standard request.

The one that resulted in me reflexively saying thank you about three times more often than usual was “to make it easy to copy and paste”.

Conversely, “must” and “absolutely” had a rework rate about twice as high as usual.

And my top-billed phrase, “make it easy to understand and consider readability,” was a case-by-case matter, with clear differences between when it worked and when it didn't.

Where is the turning point between phrases that work and those that don't?
I will deliver the answer, along with a bonus phrase to prevent garbled text.
Well then, let's see the report card for these habits.

1 | What is a mini-prompt?

When people hear “prompt,” many think of long, impressive sets of instructions.

However, what actually works in daily interactions is often just a short phrase added to the end of a request.
I have taken the liberty of calling these mini-prompts.

For example, it looks like this.

All of these are simple phrases that just slip out.
They aren't spells or advanced techniques; they are just habits.

By the way, there is a bit of defiance behind why I settled on this habit-based approach.

I see it often on the internet.

Collections of prompts that look like alphabetical spells, claiming “just enter this command to dramatically improve your answers.”

I just couldn't stand those.
You can use them if you memorize them or make a cheat sheet.
But it's a hassle to look up that cheat sheet every single time.

And I realized: when a spell works, it's not the sequence of letters that's working, but the meaning behind it.

If English-like expressions work, couldn't I just convey the same meaning intuitively in natural Japanese without needing a cheat sheet?
I'll just give up on the things I can't express in Japanese.

What remained in my hands were these habits that I could say without a dictionary.
However, because they were just habits, there was one thing I was always curious about.

Do these actually work?

Are they just “charms” I keep chanting out of assumption, or are they “tools” that stand up to empirical measurement?
Fortunately, I had the materials on hand to verify this.

2 | Verification Method — Counting My 6,233 Statements

I have almost all of my conversation logs from the beginning of this year, when I started using AI in earnest, covering about six months.

Not throwing away conversations but keeping them as assets is my treasure trove of 'thought log utilization techniques'!
This time, this treasure trove became the material for my verification.

I lined up 620 chats and 6,233 of my own statements in chronological order, had the AI aggregate them, and measured them using the following method.

1. Extract requests that included a mini-prompt.

2. Look at my own reply immediately following that.

3. If the reply was a request for a redo, such as “fix this,” “that's wrong,” or “regenerate,” count it as a failure (rework); if not, count it as a success.

In other words, instead of having the AI self-report on its own performance, this is a method of grading based on my own reactions at the time.

If I asked for a redo, it's a failure; if I didn't, it's a success. There is no judge more honest than one's past self.

I also counted the rate at which I reflexively said thank you, such as “thank you” or “perfect.”

I'm not the type to say thank you out of politeness every time, so this is not a measure of total satisfaction, but a reference indicator for the moment the output exceeded expectations.

The baseline for comparison was 1,539 requests without mini-prompts.
Their rework rate was 6.2% (93.8% success rate), and the reflexive thank-you rate was 5.1%.

Note that there are five premises for this analysis.

Please read this not as strict statistics, but asactual measurements to grasp trends.
I will use this yardstick to measure these habits.

3 | Actual results — Phrases that worked, phrases that didn't

I have summarized the results into a single report card for these habits.

There are four highlights.

🧣 First, the lowest rework rate is2.9% for "something like ~". It is less than half the benchmark and is the champion of pass rates this time (the reveal is in Chapter 7).

🧣 Second, the "spontaneous thanks" rate for "to make it easy to copy and paste" is 15.7%.to make it easy to copy and paste" is 15.7%. It is about three times the benchmark and is the champion of rates where the result exceeded expectations.

🧣 Third,"Please re-output in full" has zero rework after use. It functions as a "stop phrase" when corrections become repetitive.
AI has a habit of wanting to get by with diffs or excerpts when asked to re-output long documents. This phrase is one of the few countermeasures to negate that habit.

🧣 Fourth,the rework rate for "must" and "absolutely" is 12.0%. It is about twice the benchmark. Even though it was emphasized, the result actually worsened.

There is actually a trick to the 15.0% for "at your discretion," which looks like the bottom of the table. I will reveal this in Chapter 7 as well.

The phrase "without sugarcoating" which could not be included in the table, is a reference value because there were only 7 instances, butan interesting result appeared with a different measurement method.

When counting the rate at whichcritical words like "problems," "risks," and "issues" are heavily includedin the response when asking for an evaluation or review,it is 6.6% without "without sugarcoating" and 50% with it.

The numbers show that with one phrase, the AI's hesitation is removed.

4 | Why does "to make it easy to copy and paste" elicit thanks?

The rework rate is on par with the benchmark, butthe rate of spontaneous thanks is about three times the benchmark.
In other words, this phrase is not so much about getting a pass, but rathera phrase that brings results exceeding expectations.

I asked my AI partner directly for the reveal.

The point is the last sentence.

An effective phrase is one where the AI can judge pass/fail by itself

"To make it easy to copy and paste" is short, but the criteria for pass/fail are clear.
That is why output with a level of perfection that can be used as-is is returned consistently, and thanks come out spontaneously.

By the way, if you count requests containing "copy and paste" broadly regardless of phrasing, there were 105 cases, and the spontaneous thanks rate was 7.6%.

When narrowed down to the phrasing "to make it easy to copy and paste," the spontaneous thanks rate is 15.7%

Even with the same intent, the effectiveness changes depending on how you say it—this itself is the best proof of the idea of a mini-prompt.

The fact that the spontaneous thanks rate for "first" is surprisingly high (8.5%) is for the same reason: it becomes a verifiable condition of "dividing the work up to this point."

5 | Why don't "must" and "absolutely" work?

On the other hand, "must."
Even though I put effort into it, the rework rate was about twice the benchmark.

The reason is simple:"must" adds nothing to the content that can be executed.

AI: "'Must' works slightly as a weighting of attention, but on its own, there is zero information on how the output should be changed.

Moreover, if there are 10 'musts' in the request text, the priorities clash and everything is diluted."

Prompts that pile up tons of prohibitions and emphases have priorities that clash, making the operation unstable instead.

The barrage of "musts" is a small-scale version of that.

By the way, there is another trick to the approximately 2x rework rate.
People write "must" precisely when they are in trouble because things aren't going well.

In other words, the amount of emphasis might not be proof of effectiveness, but a sign of being cornered.

If you find that "must" is increasing in your request text, try doubting whether you can make the conditions concrete before strengthening the words.

6 | The sign "also consider readability" — Whether it works is case-by-case

Well, this is my top signature phrase.

This is my go-to phrase when requesting Word documents, and I've used it over 70 times in the last two months.
I definitely feel like it works.

Headings, font sizes, tables—it returns a readable document that has been adjusted accordingly.

However, the log numbers told a different story.
The rework rate for Word creation requests was 12.2% with 'readability' and 8.3% without.

Wait, is it worse to include it?
Breaking it down without panicking, I found three explanations.

🧣First, there is a bias where the more complex and heavy the request, the more likely I am to add 'readability'. Difficult requests have more rework regardless of whether a phrase is added or not.

🧣Second, the issue of where the effect appears. The effectiveness of this phrase shows up in the content of the Word file. Since no trace remains in the chat logs, this method cannot measure it in principle.

🧣Third, this was actually a new habit I've been using for two months. It barely appeared in logs before that. The number of trials is low, so it's too early to say anything statistically.

If you ask whether it's ineffective, the AI's view was the opposite.

Furthermore, I sometimes add parenthetical notes depending on the situation.

In short, the verdict on my signature phrase is case-by-case.

It works for outputs with room for layout adjustments, like Word or Excel.
For simple text responses, it overlaps with the default behavior, so it barely works.

Even with the same phrase, the effectiveness changes depending on the type of output.

And through this exchange, the essence of effectiveness was summarized into a single yardstick.

Effective phrase = Specificity × Scarcity (being specific and few in number)

Why 'must' doesn't work is because it has zero specificity.

Prompts that list tons of emphasis and prohibitions fall apart because there are too many important instructions, and none of them are treated as important.

'So it's easy to copy and paste' is the strongest because it adds specific pass/fail criteria in just one phrase.

I was curious, so I experimented.

I fixed the source material and the request text, and compared having four models of AI (free and paid versions as of July 2026) create Word documents by toggling only the phrase 'considering readability'.

Within the scope of my single trial for each, the results were as follows.

The free versions showed almost no reaction to slight reflection.
For the paid versions, both companies clearly changed the headings and structure, and one transformed into a design-level document with a cover page.

In other words, whether this phrase works also depends on the muscle of the model you are dealing with.

Even if the switch is the same, it won't move if there is no document design capability on the other side.
This is the second axis of the case-by-case scenario, following 'type of output'.
(Since each trial was only done once, I won't make a definitive statement, but rather treat it as a reference record.)

7 | The phrase that hands over decision-making power - 'Something like...' and 'At your discretion'

Here is the reveal for the current pass-rate champion, 'something like...'.
This was a phrase different from any of the previous types.

Here is one example from actual logs.

'Something like...' and 'Like that'

Counting the logs, the total of related expressions is about 180 times, making it a regular, actually more frequent than 'so it's easy to copy and paste'.

And the rework rate for the 34 cases used as request text was 2.9%.The phrase with the fewest redoes in this actual measurement.

The intent of this phrase is clear.

When you give specific wording like 'please use this expression,' the AI treats it as a word-for-word specification and embeds it directly into the document.
In other words, at that moment, the quality of the text is fixed to the quality of my draft.

But text generation is the AI's specialty.
The quality tends to be higher when the AI rewrites it with the same nuance rather than my own writing.

That's why I add 'something like...' to fix only the meaning and direction, and leave the wording to the AI.

Organized by type, it looks like this. "To make it easy to copy and paste" is a

phrase that adds a condition. "Always" is just a phrase that adds enthusiasm. And "something like..." is a
phrase that changes who holds the decision-making power over the wording

. However, one word of caution: if you just say "make it look good" without providing a sample, it falls into the enthusiasm category and won't work. Sample + 'something like' is the correct form

. The 'something like' phrase is effective precisely because the specificity is supplied by the sample side. This decision-making category has another sibling.

"Using your own judgment"

—this is a phrase used to eliminate the need for the AI to repeatedly ask "What would you like to do?" regarding choices that don't affect the essence of the deliverable, such as whether to use bullet points, a table, a quote, or a summary. However, those who saw the table must have noticed:


a rework rate of 15.0%, firmly in last place. There is a trick to this. If you leave the choice to the AI, the probability that the form the AI chooses will deviate from your preference structurally increases. In other words, this phrase is

a trade-off where you pay the risk of minor rework to buy back the back-and-forth of confirmation


. That is why I use it only for parts where a mismatch won't hurt. These are parts where a mismatch can be fixed with a simple "fix this" phrase. If 3 out of 10 requests stop asking for clarification, even if minor rework increases by 1, you win in the total number of exchanges. The condition for making it work is, as in the example,

to define the scope before handing it over. If you say "use your judgment" for the whole thing, the AI will start deciding even the essentials, so please be careful with that.

8 | Appendix — A Mini-Prompt Collection to Prevent Character Encoding Issues

Finally, here is a highly practical bonus. What quietly steals time when working with AI is
mojibake (garbled text). My logs repeatedly record accidents where file names were all garbled after extracting a ZIP, or batch files created by AI didn't work because of garbled text when executed. And here, the law of "effective phrases vs. ineffective phrases" appeared exactly as it is. First, let's look at an example that didn't work.

—Immediately after the request with this phrase, the file names of the delivered goods were thoroughly garbled lol

"Be mindful" is an abstract word that cannot determine pass/fail, and it is a relative of "always." What worked was
specifying the format, which I established by identifying the cause with the AI every time an accident occurred. I have placed the prescription calculated back from the accident logs here so that it is easy to copy and paste.

The key is that none of these are "don't do this," but rather a specific instruction like 'create it in this format'.

Abstract prohibitions don't work, but specific instructions do—the law of mini-prompts holds true here as well.

…And, having laid out these prescriptions, I must confess.
When I encounter garbled text, I usually just say, "It's garbled, so please fix it." It's easier that way (tee-hee).

If you're working in a back-and-forth dialogue, fix it if it's garbled is enough to get by.
However, automated processes like Python scripts that run all at once without human oversight are different.

Since no one will notice if it's garbled, this is exactly when you should include the format specifications mentioned above from the start.

9 | Summary─The True Nature of Mini-Prompts

Here is a summary of what I learned from digging through about six months of logs.

🎈 An effective phrase is a phrase that allows the AI to judge for itself whether the output is a pass or fail.
🎈 Or, a phrase that clarifies who holds the decision-making power over the wording.

Ineffective phrases are those that are just empty enthusiasm.

Looking back, all of my mini-prompts were born from actual past failures.

The experience of receiving a revised version full of omissions led to "re-output the full text," and the experience of crying over code full of placeholders led to "make it easy to copy and paste."

A mini-prompt is the minimal form of a personalized correction condition, polished from a record of failures.

Therefore, there is one method more reliable than just imitating the habits in this article.

First, just try doing this.

Next time the AI's output deviates from your expectations, don't just give up and close it; convey the specific point of deviation in one phrase and run it one more time

That extra round turns your failure into a condition, and as they accumulate, they will eventually become your own personal mini-prompts.

Habits don't grow if they are borrowed.
Only the phrases you have polished from your own failures will reliably work for you next time.

That said, please feel free to use the phrases listed here as a reference.
Give them a try if you'd like.

I hope this is helpful, and if you have any ways of using them where "this combination worked better," please let me know.

I'm sure there are people out there getting even better results with different combinations of words for the same intent (and that's my way of hedging my bets).

Finally, one bonus.
Actually, I thought I was using "make it easy to copy and paste" every single time.

When I counted, it was only a small fraction of my total requests.
Conversely, "something like..." was one of my most frequent phrases, including related expressions, even though I wasn't even aware it was a habit.

You don't know your own habits. Only the logs know.

Counting them up is actually quite interesting.

✨ If you're curious about how to keep thought logs, take a peek at my book, "Don't Throw Away Your Conversations with AI."

#AIUtilization #Prompt #ChatGPT #Claude #GenerativeAI #PromptEngineering #WorkSkills #Verification #ThoughtLog #DontThrowAwayYourConversationsWithAI #MonsieurMissClear #MiniPrompt