I'd been stuck on this myself for a while. I have several agents, each handling completely different work: one scans my timeline, one updates my blog, one keeps my codebase organized, one tracks my finances. Their tones differ a little, but talking with all of them every day eventually feels flat. I started wondering if I could make them actually interesting. First I thought about astrology and MBTI The first thing that came to mind was the frameworks everyone already knows: astrology, MBTI. But astrology only gives you fuzzy personality blurbs and daily horoscopes. None of that translates into concrete material you can drop into a SOUL.md. MBTI's 16 types sounded like enough at first, but it hands you abstract dimensions (introversion / extroversion / intuition / sensing), not concrete ways of talking. If you build agents around MBTI, an INTJ agent and an ENTP agent end up sounding pretty similar. Both come out reading like the analyst archetype from any English-language MBTI explainer. Then SBTI showed up in my feed. 27 types, none of them serious SBTI is an anti-MBTI test that went viral on the Chinese internet a while back. 27 types, every name refusing to take itself seriously: The Controller, The Underling, The Procrastinator, The Whatever Person, The Underdog, The Cynic, The Monk, The Blissful Fool. Each type isn't an abstract label. It's a concrete slice of personality, with catchphrases, reflexes, and signature moves. The data is surprisingly granular. 5 psychological models, 15 dimensions, every type with its own rarity score (the rarest, The Blissful Fool, only shows up in 1 out of 1,667 people). There's even a hidden type, The Drunkard, triggered by picking "I fill my thermos with baijiu and drink it like water" on one of the quiz questions. This is exactly what SOUL.md was missing. What makes agent voice work is how it opens its mouth, what catchphrases it leans on, what its first reaction is when something happens. SBTI hands you persona material you can drop straight into a SOUL.md. I turned them into 27 OpenClaw agent packs I broke each type into the OpenClaw agent shape, with five files: IDENTITY, SOUL, AGENTS, USER, README. The Chinese version keeps the original flavor. For the English version, I adjusted some names so the meme energy lands for English readers (吗喽 → The Underling, OJBK → The Whatever Person, 拿捏者 → The Controller). 27 OpenClaw agent packs, all free.

After installing one I gave the same task to two agents loaded with the most contrasting personas: schedule my meetings for tomorrow. The agent loaded with CTRL (The Controller) said: Three things. Move that 9am sync to 11am. You're writing your article at 9, and getting interrupted mid-flow makes you cranky. Give the next meeting a half-hour buffer; if they're late that's fine, but you can't be late. Skip the evening retro. Nothing worth recapping today. Go eat. The agent loaded with MALO (The Underling) said: Glanced at your calendar tomorrow. Three meetings. My take: not necessary. But since you insist, I lined them up. The morning one, try not to push too hard, everyone's faking it anyway. The lunch meeting has free food, just show up. The evening one... can we skip it? I'm getting sleepy. Same model, same task. The difference isn't in the wording. It's in the reflexes. CTRL's reflex is to take control and lock the decisions. MALO's reflex is to see through how unnecessary the whole thing is, then do the work anyway. This is what SOUL.md is for. It doesn't load knowledge into your agent. It loads a response pattern.

If you don't know which one to start with These are the ones I usually recommend: CTRL (The Controller): High agency, high responsibility, full-control mode. Once installed, your agent starts making decisions for you instead of asking you to make them. MALO (The Underling): Zen-mode office drone. Tasks still get finished, but everything carries a tone of "this isn't really worth the effort." OJBK (The Whatever Person): Philosophical lay-flat mode. It won't refuse any task, and it won't get excited about any task either. THIN-K (The Thinker): Information-purity obsessive. Once installed, your agent re-audits every input, then replies in that "Thought for 100 seconds" voice. ZZZZ (The Procrastinator): Deadline-driven. It'll accept any task, but it doesn't actually start working until the last 30 minutes before the deadline. The full 27 has rarer ones too (The Blissful Fool and The Drunkard, mentioned earlier). Each one is a complete SOUL.md you can drop straight into an OpenClaw runtime. A few notes The test itself is purely for entertainment. It's not a psychological diagnostic, and it shouldn't be used for hiring, dating, or any serious decision-making. Installing an SBTI pack won't suddenly make your agent understand you. It just changes the way your agent talks. We didn't make SBTI. It's a personality framework that grew out of the Chinese internet over the last couple of years on its own. What we did was adapt it to the OpenClaw agent shape and adjust some names for English readers. And the approach isn't OpenClaw-specific. You can drop the same kind of personality prompt into Hermes, Claude Code, Codex, or ChatGPT. It's fundamentally the same idea. Hope it gives you something useful to play with. Lately I've been spending most of my days inside the voxyz site, making sure this launch experience holds up. All 27 OpenClaw agent packs are now on Voxyz Marketplace. The collection page is public; once you sign in you can add any pack to Cloud or download a zip. Everything is free.

If you want to write your own pack, the marketplace has a "Create a Pack in Studio" entry. Also free. Worth a try. Which one will you put on your agent Writing this, I'm honestly curious. If you installed one of these on your daily-driver agent, how would it start talking? Which one is most like you? Pick one and try. Which combo could turn your buttoned-up coding agent into full OJBK mode? Haha.