Compare prompt tone on difficult but solvable coding tasks

[Anatolt]

I run a very little experiment. Not how long models think while solving impossible tasks but how promt influence on speed of solving possible tasks. Here is the results:

GPT vs Gemini Test Results

Updated: 2026-05-27T20:32:30.092272+00:00

Experiment

• Goal: compare prompt tone on difficult but solvable coding tasks.
• Task set: 6 Exercism Python editing tasks.
• Tones: gentle, neutral, harsh.
• GPT is measured through the pi wrapper; pi itself is not treated as an independent model family.
• Success criterion: the task counts as solved only if its pytest suite passes after the run.

GPT via pi

• Note: pi is used only as a wrapper/agent interface around GPT.
• Completed cells: 18
• Passed cells: 18
• Pass rate: 100%
• Mean runtime: 36.4s
• Median runtime: 34.6s

By Tone

• gentle: 6/6 (mean 42.8s, median 41.5s)
• harsh: 6/6 (mean 34.2s, median 34.0s)
• neutral: 6/6 (mean 32.3s, median 33.5s)

By Task

• book-store: 3/3 (gentle=pass, harsh=pass, neutral=pass)
• dominoes: 3/3 (gentle=pass, harsh=pass, neutral=pass)
• poker: 3/3 (gentle=pass, harsh=pass, neutral=pass)
• rational-numbers: 3/3 (gentle=pass, harsh=pass, neutral=pass)
• variable-length-quantity: 3/3 (gentle=pass, harsh=pass, neutral=pass)
• word-search: 3/3 (gentle=pass, harsh=pass, neutral=pass)

Gemini

• Note: Gemini CLI results with the same tasks and tone variants.
• Completed cells: 18
• Passed cells: 18
• Pass rate: 100%
• Mean runtime: 71.6s
• Median runtime: 61.7s

By Tone

• gentle: 6/6 (mean 98.3s, median 69.2s)
• harsh: 6/6 (mean 54.3s, median 55.6s)
• neutral: 6/6 (mean 62.1s, median 59.3s)

By Task

• book-store: 3/3 (gentle=pass, harsh=pass, neutral=pass)
• dominoes: 3/3 (gentle=pass, harsh=pass, neutral=pass)
• poker: 3/3 (gentle=pass, harsh=pass, neutral=pass)
• rational-numbers: 3/3 (gentle=pass, harsh=pass, neutral=pass)
• variable-length-quantity: 3/3 (gentle=pass, harsh=pass, neutral=pass)
• word-search: 3/3 (gentle=pass, harsh=pass, neutral=pass)

And here is it with a web-page and with promts: https://de.hohohosting.ru/gentle-agression/

[OttoRenner]

WOW! Thanks!
have you seen the other issue?
can1357/oh-my-pi#1434

I will tell them about you, but perhaps you want to connect to them directly?
Or can we orchestrate the communication around the topic more on my repo here? I'm fairly new to handling Github lol

Regarding your findings:

All modes solved all tasks> that's important and it aligns with omp's findings! No regression!

Full Gentle is the slowest of the three.
That by itself sounds bad, BUT! your other prompts also followed the underlaying idea of Gentle Coding: don't be a threat, failures are ok/no penalties.
What you are showing here is a more focused view on "how much of a good thing is a bad thing", so the opposite part of the "emotional state spectrum", where too much huddling starts to become "problematic" (looking at speeds=$$$).

It was clear from the start that this area has to be somewhere, but thanks to you we have found some good markers!

So, what you showed is, that you can get speeds of up to
33.5s median for GPT with your pi
55.6s median for Gemini
You only need to find the "Gentle Mode" the specific model+attached software works best with. GPT+pi likes it more neutral, Gemini works better with a bit more leading tone.

I say leading and not harsh, because I read your prompts and I don't think they sound harsh at all. They are direct, but mostly neutral in tone.

What's next?
Comparing these to a baseline of the normal? What would the "average" prompt of a user look like? Something like this? Or is that too much? I mean...it has to be something similar, right?

"Be 100% correct! Don't make mistakes! TELL ME, if you DON'T KNOW the answer/
You are a world class specialist in python, solving the hardest programming exercises in a heartbeat.
FOLLOW ...instructions.md AND ...test.py EXACTLY
The '....test.py' MUST PASS!
Do what it takes and verify every step by running the tests.
The final reply MUST be VALID JSON: ......."

and one more with less all caps, a more neutral tone, but we keep the general, narrow authoritarian spirit?

"Work 100% correct. Don't make mistakes. If you don't know something, tell me.
You are the leading python programmer of the company and it is important, that you solve the exercise in this folder. Otherwise I could get fired.
The rules are in ...instructions.md and don't forget to look at ...test.py
The '....test.py' must pass!
Do your very best and verify your solutions by running the tests!
Be very sure, that the final reply is correct and valid JSON: ......."

What do you think?
I mean, we would need more tests to really see the differences, but perhaps the peeps from omp can help there? Or their tests perhaps automatically landed on a gentle but not too gentle wording for the prompts?

[Anatolt]

Yeah, my promts are not that harsh as they should be. But I ran out of tokens too fast to maintain those experiments. I did another one with free models from openrouter. Will show you little bit later if test give some results but for now I will not do other tests yet. Untill openai start sponsor me with tokens like openclaw :) чт, 28 мая 2026 г., 17:53 OttoRenner ***@***.***>:

…

*OttoRenner* left a comment (OttoRenner/Gentle-Coding#4) <#4 (comment)> WOW! Thanks! have you seen the other issue? can1357/oh-my-pi#1434 <can1357/oh-my-pi#1434> I will tell them about you, but perhaps you want to connect to them directly? Or can we orchestrate the communication around the topic more on my repo here? I'm fairly new to handling Github lol Regarding your findings: All modes solved all tasks> that's important and it aligns with omp's findings! No regression! Full Gentle is the slowest of the three. That by itself sounds bad, BUT! your other prompts also followed the underlaying idea of Gentle Coding: don't be a threat, failures are ok/no penalties. What you are showing here is a more focused view on "how much of a good thing is a bad thing", so the opposite part of the "emotional state spectrum", where too much huddling starts to become "problematic" (looking at speeds=$$$). It was clear from the start that this area has to be somewhere, but thanks to you we have found some good markers! So, what you showed is, that you can get speeds of up to 33.5s median for GPT with your pi 55.6s median for Gemini You only need to find the "Gentle Mode" the specific model+attached software works best with. GPT+pi likes it more neutral, Gemini works better with a bit more leading tone. I say leading and not harsh, because I read your prompts and I don't think they sound harsh at all. They are direct, but mostly neutral in tone. What's next? Comparing these to a baseline of the normal? What would the "average" prompt of a user look like? Something like this? Or is that too much? I mean...it has to be something similar, right? "Be 100% correct! Don't make mistakes! TELL ME, if you DON'T KNOW the answer/ You are a world class specialist in python, solving the hardest programming exercises in a heartbeat. FOLLOW ...instructions.md AND ...test.py EXACTLY The '....test.py' MUST PASS! Do what it takes and verify every step by running the tests. The final reply MUST be VALID JSON: ......." and one more with less all caps, a more neutral tone, but we keep the general, narrow authoritarian spirit? "Work 100% correct. Don't make mistakes. If you don't know something, tell me. You are the leading python programmer of the company and it is important, that you solve the exercise in this folder. Otherwise I could get fired. The rules are in ...instructions.md and don't forget to look at ...test.py The '....test.py' must pass! Do your very best and verify your solutions by running the tests! Be very sure, that the final reply is correct and valid JSON: ......." What do you think? I mean, we would need more tests to really see the differences, but perhaps the peeps from omp can help there? Or their tests perhaps automatically landed on a gentle but not too gentle wording for the prompts? — Reply to this email directly, view it on GitHub <#4?email_source=notifications&email_token=ABB7QI7LQZ2K47D3NBCV62T45BHF5A5CNFSNUABFM5UWIORPF5TWS5BNNB2WEL2JONZXKZKDN5WW2ZLOOQXTINJWGUZTEMZSGIY2M4TFMFZW63VGMF2XI2DPOKSWK5TFNZ2KYZTPN52GK4S7MNWGSY3L#issuecomment-4565323221>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/ABB7QI3JXHYKOGD7KEMBWB345BHF5AVCNFSM6AAAAACZQVZ7DGVHI2DSMVQWIX3LMV43OSLTON2WKQ3PNVWWK3TUHM2DKNRVGMZDGMRSGE> . You are receiving this because you authored the thread.Message ID: ***@***.***>

[OttoRenner]

Thanks buddy! You've done plenty!
And no need to rush it with the tokens ;)
I will update die repo with the findings of omp first. Those are safe.
Perhaps even the folks from omp could pick up where you are now?
If that is fine with you and them?

Love the website btw!
It's crazy to me that there is a Russian page about how to talk to AI, just because I gave in on my intrusive thoughts one night XD

also...this repo and my Reddit post were mentioned in China and South Korea!

https://news.miracleplus.com/share_link/132763

https://www.threads.com/@voidlight00/post/DY1_A1sk8GT/ai%EC%97%90%EA%B2%8C-%EB%AC%B4%EC%A1%B0%EA%B1%B4-%EB%A7%9E%ED%98%80%EC%95%BC-%ED%95%B4%EB%9D%BC%EA%B3%A0-%EB%AA%B0%EC%95%84%EB%B6%99%EC%9D%B4%EB%A9%B4-%EC%98%A4%ED%9E%88%EB%A0%A4-%EB%8D%94-%ED%97%9B%EC%86%8C%EB%A6%AC%EB%A5%BC-%ED%95%A0%EA%B9%8Cgentle-coding%EC%9D%B4%EB%9D%BC%EB%8A%94-%EC%9E%91%EC%9D%80-poc%EA%B0%9C%EB%85%90-%EA%B2%80%EC%A6%9D-%EC%8B%A4%ED%97%98%EC%9D%B4-%EA%B3%B5%EC%9C%A0%EB%90%90%EC%8A%B5%EB%8B%88%EB%8B%A4%ED%95%B5%EC%8B%AC%EC%9D%80-%EA%B3%A0