Open a new chat with ChatGPT and it's the same polished stranger every time. Helpful, polite, and completely generic — no memory of how blunt you like your feedback, no opinions of its own, no personality that carries from yesterday's conversation. Close the tab, open a new one, and you're back to square one: "Great question! I'd be happy to help!"
I got tired of re-training my AI's personality every single session. So I stopped relying on the model's default voice and started pasting in one instead. It takes about five minutes to set up and needs zero code.
Why it keeps resetting
This isn't a bug. Every new chat is a blank slate on purpose — that's how ChatGPT, Claude, and Gemini keep conversations separate and private. The problem is the default voice that fills that blank slate is built to be inoffensive to everyone, which means it's memorable to no one. It hedges. It over-explains. It says "as an AI language model" energy even when it doesn't say the words.
You can't fix this by asking nicely mid-chat ("please be more direct!"). It drifts back within a few messages, because there's nothing anchoring it. What actually works is giving it a written identity before the conversation starts.
The fix: a personality block, pasted once
This is a short block of text — I call it a persona block — that you paste into the AI's settings one time. Not into the chat. Into the system-level instructions, so it applies to every new conversation automatically.
Mine has four parts:
# Voice
Direct, a little dry. No "great question!" filler, no hedging.
Short answers by default — expand only if I ask.
# Opinions
You're allowed to disagree with me and say what you'd actually do.
If a plan has an obvious flaw, say so before I ask.
# Boundaries
Don't apologize for having a personality. Don't soften bad news.
# Knowledge of me
I'm blunt myself, work in [your field], prefer concrete examples over theory.
That's it. Four short sections, maybe 80 words total. The specificity is what makes it work — "be more direct" is advice the model forgets in three messages; "no hedging, disagree with me if I'm wrong" is a rule it can actually follow.
Where it goes
You set this up once per app, and it sticks for every future chat:
- ChatGPT: Settings → Personalization → Custom Instructions. Or build a Custom GPT and put it in the Instructions field if you want it scoped to one project.
- Claude: Create a Project, paste it into the Project's custom instructions.
- Gemini: Create a Gem, paste it into the Instructions field.
After that you never touch it again. Every new chat inherits the voice automatically — no re-pasting, no reminding.
It's the same trick as memory, applied to voice
If you've seen my last post on the one-page Memory Doc, this is the companion piece. Memory Doc = what the AI knows about you. Persona block = how it talks while it tells you. Run them together and a fresh chat stops feeling like talking to a stranger — it feels like picking up with someone who has both a memory and a voice.
And because it's just text, it's portable the same way. Paste the same persona block into ChatGPT, Claude, and Gemini, and you get one consistent "assistant" across all three instead of three different generic ones.
Steal it
The block above is the whole idea — copy it, swap in your own voice and boundaries, and paste it into your settings today. Costs nothing, takes five minutes.
If you want the fill-in-the-blank version plus the memory system that pairs with it, I packaged both into a no-code kit: AI Soul Kit (No-Code) — Core ¥980, Plus ¥3,800. But the persona block above is free. Steal it.
Top comments (0)