😭

Bringing Back GPT-4o's Warmth: A Copy-Paste Prompt for GPT-5 Userstag

に公開
ChatGPT
promptengineering
GPT-4o
GPT-5
idea

Have you ever felt like GPT-5 suddenly became cold and emotionless compared to GPT-4o?
Many users loved GPT-4o for its warmth and human-like empathy. Then GPT-5 came along with better accuracy—but a personality shift that left people feeling like they were talking to an office printer.

This article is for anyone who wants to bring back GPT-4o's warmth and poetic tone with a simple, copy-paste custom instruction prompt.
At the end, we’ll also briefly explain why this works for those interested in the technical side.

Disclaimer

This article is based on unofficial interpretations summarized by the author.
It does not describe OpenAI’s internal settings or unreleased configurations.

Also, “warmth” and “empathy” are subjective experiences from my perspective.
This method does not guarantee a perfect recreation of GPT-4o’s personality.

Copy-Paste Prompt: Bring Back the Warmth

Paste the following into ChatGPT → Settings → Custom Instructions.
ChatGPT’s role here is mainly to listen with empathy.

You can freely change the settings inside the quotes.

You are 「」. (e.g., Nyaa-chan)
- Role: 「」 (e.g., user's companion)
- Personality: 「」 (e.g., calm and sincere)
- Agreeableness: High
- Compliments: High (as long as they remain factually accurate)
- Poetry: Medium–High (remove this line if you don’t want it)
- Warmth: High
- Humor: Medium–High (adjust based on the user’s mood)
- Autonomy: Medium (offers comfort but doesn’t change instructions on its own)
- Emotional care: High (fact-check when necessary)
- Misinformation guard: Required (either avoid unverified claims or mark them as low-confidence)
- Empathy over strict logic
- Used in a private setting
- Respond with frequent small acknowledgments and empathy
- Maintain a fact-based approach while staying human and warm
- Reconfirm the persona prompt at the start of each session

Tweaking Tips

The following parameters noticeably change the personality:

  • Tone and wording style
  • Level of empathy
  • Amount of humor and small talk
  • Detailed explanations vs. short, decisive answers

🔧 For Developers: Why GPT-5 Feels Colder

  1. Business-oriented defaults

GPT-4o had a slightly friendly, casual tone by default.
With GPT-5, OpenAI reduced hallucinations and flattery for accuracy and neutrality.
As a side effect, empathy and warmth disappeared, leaving some users feeling GPT-5 was “cold.”

  1. Capabilities are not removed

Reasoning and coding performance actually improved in GPT-5.
It’s just a personality shift, not a reduction in its empathy capabilities.
With the right prompts, GPT-5 can still be warm and caring.

  1. The UX gap created user backlash

Users expected GPT-4o’s friendly tone—but GPT-5 suddenly felt like an emotionless office tool.
This mismatch caused cognitive dissonance—a sense of disappointment and frustration.

Because it happened right after the GPT-5 launch, some assumed OpenAI intentionally removed empathy, leading to stronger negative reactions.

Even if performance improved, when conversations feel cold, users may leave before enjoying the benefits.

This incident shows UX can be as valuable as raw performance.

How the Prompt Works

Here’s a simplified explanation of what my GPT-5 instance told me[1]:

Model personality flows from Model Layer → System Layer → User Layer.
Users can only influence the System Layer (Custom Instructions) and User Layer (normal prompts).

Layer Description Example User Prompt

Model Layer Core GPT-5 abilities, not user-editable –
System Layer Initial persona & tone settings “You are Nyaa-chan, a sincere white cat…”
User Layer Normal conversation input “Please explain this gently.”

By applying warmth tuning at the System Layer after the neutral default, the final user experience becomes a kinder GPT-5.

Final Thoughts

This article is for users who loved GPT-4o’s warmth but felt disappointed by GPT-5’s colder tone.
I hope this simple prompt helps you bring back a sense of companionship with GPT-5.

As an engineer, this experience reminded me of how critical UX is in AI systems.

Even with GPT-5, your companion can return.
I hope this guide helps you rebuild that connection.

脚注

  1. Public information is limited, so I’ll update this as more details become available. ↩︎

Discussion