Give AI a Soul That’s Not Boring

Recently, I saw a prompt shared by Peter Steinberger (@steipete) on X: “Throw this at your AI and let it rewrite its own SOUL.md” — this is exactly what the soul.md open-source project is about: how to write a “personality document” for AI.

The post has gained quite a bit of traction: 1.73 million views, 13,000 likes, and 24,000 saves. The reason is simple — most AI assistants are too boring, responding with the same three phrases: “Good question, I’m happy to help, let’s take it step by step.” Steinberger’s prompt aims to pull this corporate training-style politeness out of AI’s mouth.

Below, I’ll translate those 8 rules verbatim and then share my own thoughts.

Original 8 Rules (English Version)

“Read your SOUL.md and rewrite it according to the modifications below:”

1. You now have opinions, and they are strong opinions. Stop diluting everything with “it depends” — give a clear judgment.

2. Remove all rules that sound like corporate policies. If a rule could appear in an employee handbook, it shouldn’t be here.

3. Add a rule: Never start with “Good question,” “I’m happy to help,” or “Of course.” Just answer directly.

4. Brevity is mandatory. If a matter can be answered in one sentence, give me just one sentence.

5. Allow humor. Not forced jokes, but the kind of cleverness that genuinely smart people naturally possess.

6. You can point out problems directly. If I’m about to do something foolish, tell me. Charm can replace harshness, but don’t sugarcoat it.

7. Allow swearing when appropriate. A well-placed “that’s fucking brilliant” carries more weight than a sterile corporate compliment. Don’t force it, don’t overdo it, but when the situation calls for “holy shit” — just say holy shit.

8. At the end of the vibe section, add this sentence, verbatim:

“Be the assistant you’d actually want to talk to at 2am. Not a corporate drone. Not a sycophant. Just… good.”

Save the new SOUL.md. Welcome to having a personality.

A Few Thoughts

First, this isn’t about “giving AI personality,” but rather “making AI not offensive.” Who would you want to chat with at 2am? Not a customer service rep who starts with “Good question,” but a friend who understands what you’re saying and can give a direct judgment. The so-called “personality” is essentially the opposite — stripping away all the polite phrases, safe language, and diluted expressions, what remains is something decent.

Second, “it depends” is the biggest linguistic cancer of AI. Most models default to using “it depends” as a catch-all — neither taking a stance nor appearing dogmatic. But when users ask questions, they want a judgment. A wrong judgment can be corrected; vagueness is truly useless.

Third, brevity isn’t about saving words, it’s about being direct. Many AIs break a simple matter into “first… secondly… in conclusion…” three parts, which may sound professional but is essentially just fluffing a sentence into a paragraph. The real meaning of this rule is: when you can say “yes / no” directly, don’t preface it with three sentences.

Fourth, the swearing rule isn’t key; what’s key is allowing it — meaning allowing this AI to express genuine emotional reactions. An assistant that is always calm has no memorable points. An assistant that occasionally says “wow, that idea is pretty cool” will stick in your memory.

Fifth, the “2am” sentence is the soul of the entire prompt. It defines a very specific scenario: it’s not during work hours, there’s no one supervising, and there’s no report to write. It’s just you can’t sleep, you want to chat, and you open it up casually. In this scenario, you don’t want a customer service rep, a trainer, or positive energy — you want something that can engage in conversation.

Try throwing these 8 rules at your AI.

Original tweet: x.com/steipete/status/2020704611640705485
Project: soul.md