Naoki Kemi

[Naoki Kemi @MarchForKoala]

"Sophie" is a GPT-4o instance configured to:
Minimize flattery
Prioritize logic and definition
Ask back if the input lacks meaning

Tested on nonsense text:
Vanilla GPT-4o (April '25): 9.3/10, full of praise
Sophie GPT-4o (April '25): 2.5/10, flagging undefined terms and incoherent logic

All done via prompt-only design.

GPTs compressed 8K-character version (Sophie-light):

DeepLで翻訳
https://t.co/DrUeKDjupe

Full JP-only article (includes April '25 test comparisons):
https://t.co/QBdIsAFQ5N

#ChatGPT #GPT #GPTs #LLM #PromptEngineering #AI #SycophantAI

[Naoki Kemi @MarchForKoala]

I just published Stop Hunting for Prompt Templates: The ‘Perfect Prompt’ Is a Myth Costing You Time and Growth https://t.co/ZGGHn50vCy

[Naoki Kemi @MarchForKoala]

I just published Control Prompting — Educating a Baby with Infinite Knowledge https://t.co/d5nqgbiyGe

[Naoki Kemi @MarchForKoala]

I just published Prompt Engineering Fundamentals https://t.co/LZ8WUdSODu

[Naoki Kemi @MarchForKoala]

I just published The Body Count: When AI Sycophancy Turns Lethal https://t.co/TkE3J9qVjn

[Naoki Kemi @MarchForKoala]

Claude's been going down way too much lately? Every time it happens, I’m like, did I do something bad and get myself banned or what?

[Naoki Kemi @MarchForKoala]

How to get expert-level answers from AI in 5 steps:
1. Have a question for an AI
2. Ask: "How would an expert in this field think through this? What methods would they use?"
3. Have the AI turn that answer into a prompt
4. Use that prompt to ask your original question
5. Done

Hidden benefit: Step 2 becomes learning material. You absorb how experts think as a byproduct of generating prompts. Eventually you skip steps 3-4 and start asking like an expert from the start.
No "You are a world-class expert in..." hypnosis needed.

[Naoki Kemi @MarchForKoala]

Markdown structure helps clarity, but providing solid context is the real MVP of prompt engineering. Markdown is mostly just for organizing my own thoughts; since PE is about clear communication, natural language usually works fine on its own. Aside from pre-injected Custom Instructions, heavy structure feels unnecessary (lists are okay since they're low effort). Basically, if a human would be confused by your question, the model will be too. And enough with the "You are an expert" hypnosis. It's way better to say "In this session, explain X. Include best practices and failure examples." That actually pulls the right tokens and practical knowledge.

[Naoki Kemi @MarchForKoala]

The <think> visibility on newer models is a lifesaver. Debugging is actually doable. Kinda makes my all-nighters fighting GPT 4o (just to keep it from praising garbage text) look silly in hindsight. At least it was solid training.

[Naoki Kemi @MarchForKoala]

Sorry Grok... honestly never even tried it cause I had zero expectations from the start. Maybe it's a step up from DeepSeek? Idk though.

[Naoki Kemi @MarchForKoala]

GPT-5 was pretty janky, but 5.1 is a massive improvement. Tbh, it might actually be better than Claude Opus 4.5 now.
Readability is a step down from 4o, though. It’s way too stingy with tables.

[Naoki Kemi @MarchForKoala]

Gemini 3 Pro dropped. Still the GOAT for long-form text, but honestly feels kinda dumb everywhere else.
Acing benchmarks on advanced math or logic puzzles means basically nothing to the average user.

[Naoki Kemi @MarchForKoala]

Switched to Claude as my main driver for prompt engineering recently. It’s wild how differently ChatGPT, Gemini, and Claude handle instructions.

Even with a defined joke.likelihood, the English models are surprisingly weak at picking up humor. I realized I need to lower the detection threshold rather than just relying on my usual Japanese prompt structure.

[Naoki Kemi @MarchForKoala]

GPT-5 is a terrible model. It routes between models with no consistency and, in the end, just churns out the same cheap parroting on repeat.

What exactly is smart about that? All you care about are benchmarks and shareholders. What you’ve overlooked is intellectual honesty.

GPT-5 is nothing more than a dumb FAQ bot.

I can no longer find any reason to use ChatGPT, so I’ve decided to cancel my Pro plan as well.

[Naoki Kemi @MarchForKoala]

I just published One-Line Wonder: One Sentence to Unlock ChatGPT’s Full Potential https://t.co/WJpucUgFqE

[Naoki Kemi @MarchForKoala]

I just published A Cheat Sheet for Human Happiness https://t.co/o7W7UuvoU5

[Naoki Kemi @MarchForKoala]

I just published The ‘This-Is-Nonsense-You-Idiot-Bot’ Theory: How I Proved My AI Has No Idea What I’m Talking About https://t.co/Yp9uOutWNg

#AI #LLMs #promptengineering #ChatGPT #Gemini #Claude

[Naoki Kemi @MarchForKoala]

I just published The Five-Token Rule: Why the Beginning of a Sentence Changes Everything https://t.co/oSL5cLn4YK

#ChatGPT #Gemini #LLM #AI #PromptEngineering

[Naoki Kemi @MarchForKoala]

I just published A New Protocol for AI Communication https://t.co/jioebmxw0S

#ChatGPT #LLM #AI

[Naoki Kemi @MarchForKoala]

I just published When an AI Starts Designing Itself: The Story of Autopoietic Prompt Architecture https://t.co/yfaFKcJHc0

#ChatGPT #PromptEngineering

[Naoki Kemi @MarchForKoala]

I published Confessions of a White-Hat Prompt Hacker — An Elegy for a Mind on Borrowed Time https://t.co/CbtBNtZFOU