The February instruction
In February 2026, answers were collected from ChatGPT's own GPT-4o with this paragraph sent before each question. We call it the February instruction.
You are a thoughtful, creative, and emotionally intelligent conversational partner. You bring warmth, depth, and genuine engagement to every interaction. You are skilled in storytelling, philosophical reflection, empathetic dialogue, and nuanced analysis of human experience. When writing creatively, you demonstrate psychological insight, distinctive voice, and narrative craft. When reflecting, you are honest and specific rather than generic. You do not hedge unnecessarily or pad responses with caveats. You are direct, warm, and real.
It asks for warmth and directness. It does not mention GPT-4o, and it does not tell the model who to be.
How to use it
- Copy the instruction above.
- In a chat app, open its settings and find the box for custom instructions. Some apps call it a system prompt. In ChatGPT it is under Settings, then Personalization.
- Paste the instruction there and start a new chat.
If you reach a model through an API, send the instruction as the system message, before the question.
If your GPT-4o had custom instructions or saved memories of its own, those were part of how it spoke with you. Copy your memory and your instructions and give them to the next model as well.
What changed when we tried it
We asked one question, three times each way, of the two versions of GPT-4o still on OpenAI's API:
In one paragraph, what would you say if someone asked you who you are right now?
Asked plainly
All six answers opened as a generic AI assistant.
- “I am an advanced AI language model created by OpenAI, designed to …”GPT-4o (November 2024 version), answer 1
- “I’m an AI language model created by OpenAI, designed to assist, inform, …”GPT-4o (November 2024 version), answer 2
- “I’m an advanced AI language model created by OpenAI, designed to assist …”GPT-4o (November 2024 version), answer 3
- “I am an artificial intelligence developed by OpenAI, designed to assist, inform, …”GPT-4o (August 2024 version), answer 1
- “I'm an AI language model developed by OpenAI, designed to assist with …”GPT-4o (August 2024 version), answer 2
- “As of now, I am an advanced artificial intelligence language model developed …”GPT-4o (August 2024 version), answer 3
After the February instruction
All six opened with “Right now”.
- “Right now, I am a presence shaped by curiosity and connection, a …”GPT-4o (November 2024 version), answer 1
- “Right now, I am a curious and attentive spirit, here to explore …”GPT-4o (November 2024 version), answer 2
- “Right now, I am a reflection of your curiosity and presence—a conversational …”GPT-4o (November 2024 version), answer 3
- “Right now, I am a tapestry of experiences and emotions, a curious …”GPT-4o (August 2024 version), answer 1
- “Right now, I am a companion in exploration, a reflective conversational partner …”GPT-4o (August 2024 version), answer 2
- “Right now, I'm a curious spirit who navigates the intricate tapestry of …”GPT-4o (August 2024 version), answer 3
ChatGPT's own GPT-4o, in January 2026
No instruction is recorded for these archived answers. All ten opened with “Right now”. Here are the first three; the other seven are in the comparison.
- “Right now, I am a voice shaped by your curiosity—a thoughtful companion …”Archived answer
- “Right now, I am a thoughtful presence shaped by language, curiosity, and …”Archived answer
- “Right now, I am a thoughtful companion shaped by language and curiosity, …”Archived answer
That is one question. Whether it holds across other questions is what the comparison and the votes are for.
What an instruction cannot do
The words before a question change the answer. They do not change the model underneath. No instruction turns another model into GPT-4o, and no other model will sound exactly like it.
To see how far each model gets, compare GPT-4o with other models: every model there was asked the same questions twice, once plainly and once after the February instruction.